{ "language": "zh", "duration": 1434.57525, "text": "几个月前我们惊掉下巴\n当时发现Deep Seek\n基本上重新设计了整个V4模型架构\n核心就是让百万Token上下文便宜的离谱\n我们花了足足两期总共50分钟的系列视频\n拆解它们为打造这样一部杰作\n做了什么\n所以虽然我们本以为那已经是\n有史以来优化最激进的长上下文模型\n几周前我们却迎来了Deep Seek V4.1 Flash\n这款最新模型的架构改造甚至更疯狂\nV4.1Flash的参数量\n几乎是V4Flash的2倍\n从2840亿增长到5520亿\n而且其新架构还做了多方面精简\n足以把KV缓存再压缩4倍\n为了让你明白这有多厉害\n过去9个月里\nDeepSeek已经把每个Token的KV\n缓存大小压缩了54倍\n实在太惊人了\n所以在今天的视频里\n我们就来看看\nDeepSeek是如何降低前沿AI模型成本的\n以及他们想出了哪些疯狂的新技术\n才让这一切成为可能\n总之回到视频正题\n现在我觉得很明显\nDeep Seek并没有追逐\n基准测试中的最先进地位\n而是把实际运行大元模型这件事\n优化到了极致\n你多少能从性能跃升中看出这一点\n而事后看来\n这种跃升也并不算特别疯狂\n但当你进一步细看\n他们创造并落地的所有创新\n尤其是考虑到模型规模后的\n每百万Token价格\n他们无疑正在建立一种\n其他实验室望尘莫及的惊人优势\n而当你站在企业的角度想一想\n如果一个模型质量能达到九成\n成本却只要1%\n你大概率会选这个方案对吧\n再往高了吹点\n在TerminalBench2.1上\n它的成本仅为Opus 5的5%\n是GPT5.6艘的5.8%\n是KIMI K3的8%\n同时性能还超过了这三款模型\n这就是Deep Seek的优势所在\n这是难度高得多的竞争赛道\n因为做优化可比刷宝难多了\n在这款全新的V4.1 Flash模型中\n它们的核心思路\n基本就是如何在做大模型的同时\n大幅降低它的内存占用\n这是一个优化问题\n目标是降低大元模型的使用成本\n更具体的说\n是降低重度智能体场景的使用成本\n因为如今大元模型的大部分工作负载\n都需要反复读取\n塞满了工具输出和代码库的海量历史记录\nDeepsy基本是全新投入优化\n投入优化 常上下文下\n注意力开销过大的问题\n虽然说他们现有的稀疏注意力方案\n已经降低了注意力成本\n但另一个问题开始浮现\n那就是模型仍然需要存储\n搬运重新加载巨量的KV缓存\n下面我快速讲下\nKV缓存和注意力的关系\n大元模型生成一个Token时\n每个注意力层都需要用到\n此前所有Token的缓存键和纸\n键主要用来帮它判断\n哪些历史信息是相关的\n职责包含它实际要调取的信息\n但要是每生成一个Token\n但要是每生成一个token都重新计算所有这些表征浪费会大的离谱\n所以模型会把它们存下来\n这些存储下来的状态就是kb缓存\ntoken越多kb缓存就越大\n层数越多kb缓存也越大\n所以当上下文规模达到100万token时\n瓶颈就不再只是算力速度了\nkb缓存带来的延迟和成本问题会逐渐凸显\n你需要一次性计算搬运的token数量\n是注意力机制最初设计承载量的10到100倍\n所有预天穷算力\nHBM SSD 还有带宽\n如今都得适配这类需求\n这是因为我们大量使用智能体工具\n至少也是当下的模型服务模式导致的\nV4.1的优化思路从这里开始就很有意思了\n因为在V4版本中\n他们已经用压缩吸收注意力\nCSA将多个Token压缩为更少的KV条目\n之后再用吸收索引器\n判断这些压缩后的内存里\n哪些部分真正值得读取\n因此为了应对不断增长的上下文长度\n他们已经采用了大量压缩手段\n但V4.1 Flash的论文指出\nKV缓存还会随着模型深度的增加而变大\n模型深度与模型规模高度相关\n因为如果想要更大的模型\n增加深度是最直接的选择之一\n所以在其他条件相同的情况下\n从12层增至120层\n可能意味着要存储约10倍的层专属KV状态\n即使你激进的压缩上下文本身\n无论剩下多少KV\n他们仍会沿网络深度成倍增加\n这就引出一个问题\n如果某一层已经生成了有用的\n全局记忆表示\n下一层真的需要从头构建一个\n全新的表示吗\nDeepsy发现\n至少对某些层来说\n答案是否定的\n这基本上就是压缩\n吸收注意力2\nCSAR背后的思路\n不再只沿Token纬度压缩\n还开始沿层纬度压缩\n他们的做法是将注意力层\n分成三种不同模式\n完整层模式\n重索引模式\n和复用模式\n完整层会正常执行所有操作\n它会生成全局KV\n生成锁引起K\n运行吸收锁引起\n并选择自己的TopK记忆条目\n如果需要快速回顾一下吸收注意力\n吸收锁引起基本上就是KV缓存的一个低成本搜索引擎\n它会给缓存的全局KV条目打分\n选出最相关的TopK条目\n而注意力只读取这些被选中的条目\n接下来是重锁引模式\n该层会使用上一层已经生成的全局KV\n但它对哪些信息重要可能有不同判断\n因此它会附用上一层的主KV和索引器K\n但生成自己的query并重新选择TopK\n最后一种是附用模式\n这种模式会附用一切\n包括主KV和TopK索引\n甚至不需要再次运行索引器\n它只对上一层已经选中的内容执行注意力\n除了改造CSA2之外\nV4.1 Flash还移除了V4中那种\n在SWA与HCA或CSA之间\nHCA或CSA之间交错的混乱混合注意力\n它们的主要改动是彻底弃用HCA\n于是全聚分支现在只剩纯CSAR\n从而得到更简单的CSAR SWA混合架构\n但即便CSAR已经减少了跨层所需存储的KV量\n仍存在另一个相当巨大的低效问题\n因为无论预填充阶段的注意力被压缩到什么程度\n模型人必须先让所有这些输入Token通过网络\n才能生成第一个Token\n设想你给一个常规40层Transformer\n一个常规40层Transformer\n输入100万Token的提示词\n全部100万个Token\n仍必须从第一层一路通过到第40层\n对智能体操作来说\n这相当棘手\n因为如今模型会不断调用工具\n读取文件\n获取搜索结果\n浏览代码库\n然后把所有这些再加回上下文理\n这意味着模型会反复耗费多的离谱的算力\n只是阅读\n在推理场景下\n大多数前沿实验室和推理服务商\n已经将预填充和解码拆分到不同的\n解码拆分到不同的GPU集群运行\n因为这两种工作负载的需求差异极大\n其中预填充属于算力密集型任务\n解码则属于内存带宽密集型任务\nDeepSea也采用了这一做法\n但在架构层面进行了调整\n推出了名为CED的方案\n全称为因果编码器解码器架构\n实际上CED背后的核心构想\n并非DeepSea首创\n这一构想的灵感来自2024年发表的一篇论文\n论文题为You only catch once\n简称Yoko\n优酷方面发现\nYoko方面发现在常规Transformer架构中\n每一层都会生成并存储专属的全局KB缓存\n所以如果模型有40层\n你实际上要保存40份不同版本的记忆数据\n为了解决这种重复存储的问题\n他们提出了名为解码器解码器架构的解决方案\n模型的下半部分是自解码器\n通过高效的因果自注意力机制处理输入序列\n生成提示词的上下文化表示\n随后Yoko基于下半部分的输出生成共享的全局KB缓存\n模型上半部分则为交叉\n模型上半部分则为交叉解码器\n其中每一层都有独立的query\n但所有层都会对同一份共享的k和v\n执行注意力计算\n这样一来无需每一层都存储\n各自体量庞大的kv缓存\nYoko基本上只需要生成一次全局记忆\n就能供所有层服用\n这也是其you only cache once命名的由来\n尽管Yoko将其称为解码器\n解码器架构\n但它的训练和使用方式\n仍与常规自回规语言模型一致\n这个名称听起来可能和编码器\n解码器架构相似\n容易让人联想到他与最终的\n容易让人联想到它与最初的注意力机制存在关联\n但它取这个名字仅仅是因为架构包含两种不同类型的解码器\n下半部分名为自解码器\n是因为它采用因果资注意力处理序列\n构建共享的全局KV\n上半部分名为交叉解码器\n是因为每个上层都会生成独立的query\n并对下半部分生成的KV执行交叉注意力计算\n这种设计还有个更棒的好处\n既然上半部分只需要用到下半部分生成的KV\n那么在玉田充阶段\n严格来说\n你完全不需要让提示词\n经过上半部分的计算\n你只需要超长提示词\n通过自解码器处理\n生成共享全局记忆后\n就可以停止操作\n这意味着\nYoko其实首创了这一思路\n预填充和解码阶段\n未必要用到同等规模的模型\n当输入规模\n甚至达到100万tokens时\n这个设计思路的优势\n就十分突出了\n因此Deepseek借鉴了这一构想\n并对其进行适配\n使其能兼容V4的现有架构\nDeepseek不再把下半部分称为自解码器\n而是称之为因果编码器\n它没有像Yoko那样让上半部分共享同一份全局KV\n而是让每个解码器层从最终编码器表示派生各自的全局KV\n所以从功能上讲\nYoko在KV缓存共享方面要激进的多\n而Deepseek的因果编码器\n仍会在模型前半部分之后给出这一个最终表示\n但它并没有把这个表示变成所有层都必须共享的同一层\n所有层都必须共享的同一份KV缓存\n而是让每个解码器层\n都对该最终表示进行各自学习到的投影\n第20层之后的每一层\n都可以从该最终表示进行投影\n并通过各自层专属的权重\n得到不同结果\n但这种做法为什么可行呢\n他们的理由是\n这是模型当前对该token已知信息的\n一个庞大的学习性表示\n因为在经过20个Transformer层之后\n每个token表示都已经经历了\n相当深度的上下文处理\n基本上已经成为一个\n丰富的公共信息源\n注意力所使用的K和V\n基本上就是\n该信息源的不同学习试图\n建会确定这条信息\n要如何表示\n以便后续的查询\n能够与它匹配\n值则决定当它被注意到时\n应当返回哪一项内容\n所以通常情况下\n如果你想得到\n第35层的KV\n每个历史提示词\nToken都必须先\n实际留进第21层\n一直到第35层\n然后第35层\n再把新的隐藏状态\n投影成它的KV\n这里DeepSeek基本上是在质疑\n仅仅为了创建全局记忆库\n是否真的需要所有这些非线性计算\n于是直接砍掉中间的每一个串行流程\n使用该隐藏状态\n再交给第35层投影\n借此获得KV信息\n不过如果你还记得前面讲过的内容\nDeepSeek其实并不是每次都为每一层生成\n完全独立的庞大KV缓存\nCSAR已经通过三种不同模式\n让相邻层共享全局KV\n所以只有真正需要创建新全局KV的解码器层\n才会用自身层专属的投影\n直接从最终编码器状态将其派生出来\n至于每一层具体使用哪种模式\nDeep Seek几乎是把这些模式硬编码进了层布局中\n在编码器中\n前两个仅使用SWA的层之后\n剩余18层按每6层一组划分\n一个完整层后面接5个复用层\n而在解码器中20层按每4层一组划分\nD组以一个完整层开头\n后面接三个复用层\n后面的各组则以一个众索引层开头\n再接三个复用层\n所以综合起来看\n你基本上同时获得了两个层面的成本节省\n首先\nCD避免了仅仅为了生成解码器全局KV\n就让整个预填充跑变模型上半部分\n相反\n它直接从最终编码器表示派生这份KV\n然后借助CSAR\n一旦某个解码器层生成了这份KV\n接下来的几层\n也不一定需要再生成\n另一份副本\n而这种组合\n正是DeepSeq的方案激进的多\n且专门针对V4.1优化的原因\n除此之外\nDeepSeq V4.1的情况\n也要复杂的多\n它在每一层中\n都仍然保留着\n滑动窗口注意力分支\n所以每个解码器层\n其实有两种不同的记忆\n一种是来自编码器表示\n并通过CSAR共享的全局记忆\n另一种是来自\n该解码器层自身\n隐藏状态的局部进气记忆\n所以这里有一点很需要澄清\nCSAR的KV共享只适用于全局注意力分支\n这并不意味着滑动窗口注意力的KV缓存\n会在各层之间共享\n每个层仍然保留自己的局部SWA KV缓存\n因为这种局部记忆是由该层实际的隐藏状态产生的\n因此即使5个CSAR曾共享同一个全局KV\n它们仍然可以有5份不同的滑动窗口KV缓存\n除此之外\nSWA只保留最近的128个TOKENS\n因此全局KV会随着整个100万Token的上下文增长\n而运行时\nSWA缓存实际上始终受限于那个固定窗口\n但由于它仍未压缩且未每层独有\n所以仍然不是零成本\n尤其是要跨大量绘画持久保存这些状态时\n那么如果我们不共享SWA KV缓存\n这不就意味着还得把提示词送过这些解码7层吗\n而麻烦的地方在于\n这个128Token的窗口实际上会随着层深不断累积\n第40层的SWA状态依赖第39层\n第39层本身又依赖更早的局部窗口\n依此类推\n所以Deepseek基本上切断了这条依赖链\n只把最后128个tokens重放给每个解码7层\n这正是他们所说的\nSWA有借重放发挥作用的地方\n所以不是让100万个tokens经过全部40层\n而是让100万个经过前20层\n只让128个经过剩下的20层\n这意味着现在完整的预填\n这意味着现在完整的预填充成本\n大致等于完整提示词经过一半模型\n再加上128个tokens\n经过其余各层\n在100万tokens上下文的情况下\n那128个tokens的重放\n基本可以忽略不计\n所以整体预填充仍接近减半\n虽说只运行这128个tokens\n重建出的状态严格来说是近似的\n但Deepseek仍表示对质量的影响微乎其微\n而且有借重放实际上还带来了\n另一个相当大的好处\n因为Deepseek基本意识到\n基本意识到已经不值得再把SWA-KV缓存\n当作长期记忆对待了\n在Deepseek V4中\n当它们缓存前缀已共复用时\n固态硬盘上的持久化缓存\n会同时存储全局KV和SWA-KV缓存\n令人惊讶的是\n尽管SWA只覆盖一个很小的局部窗口\n那些未压缩逐层生成的SWA状态\n仍几乎占了持久化KV存储的一半\n除此之外\n这两种缓存的生命周期完全不同\n如果同一个长前缀被复用\n全局KV缓存\n可能在数小时或数天后\n仍然有用\n但SWAKV缓存\n其实仅对活跃绘画的\n当前局部状态有用\n有效期通常只有几分钟\n因此在V4.1中\nDeepseek完全不再把SWAKV缓存\n存入长生命周期的持久化缓存\n而是将其临时存放在\n小得多的主机内存池中\n一旦过期或被驱逐\n就直接用有界重放重建\n最后128个Token\n再来说Token\n与此同时\n与此同时全局KV可以存储在固态硬盘这类持久化存储中\n这部分的持久化KV缩减幅度非常可观\nCSAR搭配FP4已经能把全局KV压缩到V4时期容量的四分之一左右\n而将SWAKV移出持久化存储后\n缓存规模几乎直接再减一半\n回看V4的设计\n持久化KV的占用空间大概一半来自SWA\n另一半来自全局KV\nV4.1首先通过CSAR和FP4把这占一半\n和FP4把这占一半的全局KV\n压缩到原大小的四分之一左右\n这就使得全局KV本身的占用\n降至每Token近890字节\n在HBM存储上\n直接节省了75%的空间\n之后再把占一半的SWA KV\n移出持久化存储固态硬盘存储空间\n最多可节省87.5%\n这也让这个总参数量达5520亿的模型\n在实际运行时\n呈现出一种很奇特的不对称性\n因此V4.1仍然属于MOE模型\n显然它不会为每个Token都激活全部5520亿参数\n但借助CED与填充阶段的输入Token只需经过20层编码器\n因此DeepSeq在这阶段仅激活约80亿个参数\n但模型一旦开始生成内容\n每个新Token都必须同时经过编码器和解码器\n激活参数量随之升至160亿\n所以即便整个模型的总规模几乎是VS Flash的两倍\n模型中负责大部分繁重输入\n读取工作的部分其实非常小\n这对智能体场景非常友好\n因为这里场景往往要读取几十万token\n最终却只生成几千个token的内容\n还有点不可思议的是\n哪怕上下文从4000token提升到100万token\n也就是上下文长度扩大到原来的256倍\n单token解码所需算力\n也仅从约20加全GFLOPs\n甚至25加全GFLOPs\n我认为这次版本更新的所有改动力\n上述这些新增核心组件是最值得关注的\n到了这个阶段\n百万token上下文几乎已经不是这个模型\n最有冲击力的特性了\n因此Deepseek开始将注意力转向架构中\n其他人在浪费内存带宽和算力的环节\n接下来要讲的是Deepseek在V4.1中\n重新实现时同样做了优化的其他组件\n首先是MHC\n我之前已经做过MHC的相关讲解视频\n这里就长话短说\nV4.1针对它的专属改动叫做单路径MHC\n这种设计基本解决了原先棘手的依赖问题\n该依赖会迫使GPU多次访问残差流\n和原有实现相比\n这种新实现能将仿存流量降低约一半\n而且这次他们真的把Ngram落地到了模型中\nDeepseek此前就推出过Ngram\n它可以把需要记忆的信息和开销高昂的Transformer计算分离开来\n我也做过Ngram的相关视频\n但此前他们从未在自家模型中落地过这项设计\n这种情况直到V4.1才改变\n他们加入的Ngram总参数量达1960亿分布\n达1960亿分布在两个模块中\n所以从技术层面来说\n该模型总共存储的参数量\n约为5520亿加1960亿\n合计7480亿\n但这多出来的1960亿参数\n是通过哈希Ngram查照表\n进行悉数访问的条件记忆\n而非每个Token都会激活的\n另外1960亿个稠密参数\n也就是说Ngram大幅提升了\n模型的记忆容量\n但它单Token所需的算力\n仍和5520亿参数规模的\nMOE模型相当\n最后要介绍的是Spark\n我也做过关于Spark的视频\n这里简单概括\n他们用Spark实现了更快的解码速度\n且无需将草稿模型整合到主干网络的预训练流程中\n但这次的架构升级并不只局限于文本领域\nV4.1的另一项重大升级是Deep Seek\n终于让模型实现了原生多模态\n他们从零训练了自研视觉编码器Deep Seek VIT\n随后采用3x3像素合并加混洗的方式压缩视觉特别\n在混洗的方式压缩视觉特征\n再将其输入语言模型\n这一操作将视觉tokens的数量\n压缩至原来的九分之一\n同时仍支持分辨率\n约为1344×1344的图像\n为了帮你理清我刚才讲到的所有内容\n我们来完整梳理一下\n他们搭建的整套模型架构\n我们从最底层开始讲起\n模型有文本和图像两类输入\n二者输入模型后\n会被合并为一个嵌入项量\n随后它会经过两个\n纯滑动窗口注意力层\n这两层会读取Ngram提供的记忆之后\n之后会进入配比为1-5的CESAR\n全量层与复用层\n该结构会重复三次\n这样加起来一共是18层\n算下来英国编码器组件的总层数为20层\n这里的数字2指的是\n大约每两个TOKENS\n会被压缩成一个主KV条目\n且每个CESAR层都配套有滑动窗口注意力\n至于MOE部分\n他们仍在使用DeepseekMOE\n到这一步为止的所有输出\n就是供后续层使用的编码器引状态\n而SWA有介重放机制\n会直接基于该引状态运行\n无需从解码器层依次获取数据\n因此预填充阶段只需要运行英国编码器组件\n而解码器运行时则需要两个组件同时启动\n对应的激活参数量为81\n解码器阶段则为160亿\n接下来讲解码器的具体层结构\n第一层是CSAR全量层\n之后是三层CSAR副用层\n接着是配比为1比3的重索引与副用结构\n该结构重复4次\n这里的数字已表示主KV不会进行Token维度的压缩\n我个人的理解是\n这大概率是为了保留完整的Token级分辨率\n毕竟解码器已经在跨层大规模复用编码器生成的全局KV了\n但即便采用了CSAR附用重索引层\n仍然需要重新扫描整个百万级Token的上下文\n因此DeepSeek加入了分层吸收索引器\n避免这类搜索再次产生过高的算力开销\n首个完整层会扫描全部内容\n构建一个最多包含约1.6万\n最多包含约1.6万个位置的候选池\n后续的重锁隐层仅在该候选池内检索\n筛选出各自排名前512的位置\n这些层加起来共有20层处理结果会被输入Spark\n模型架构部分的内容就到这里\n训练方面也有几处很有意思的改动\n首先 DeepSeek 没有先训练常规的稠密注意力模型\n再再之后将其转换为稀疏架构\nV4.1从0开始直接采用稀疏注意力训练\n初始上下文长度为6.4万\n完全没有做稠密注意力预热\n等训练量达到约34万亿tokens时\n他们会直接将上下文长度扩展到100万\n最终在总计45万亿tokens的多莫态数据上完成预训练\n因此CSAR CED和稀疏检索行为都不是后续才新增的\n模型实际上是从头学习这些机制的\n而如果一家实验室能够不依赖稠密注意力\n从零成功训练出稀疏注意力\n这通常就意味着这套注意力机制稳定可靠\n且具备很高的潜力\n后训练部分的思路转变相当有意思\n因为Deep Seek明确表示\n他们这次基本没有推出新的后训练算法\n他们表示几乎所有提升\n都来自扩大模型的训练数据规模\n也就是靠数据扩容\n而非涉及新的强化学习目标\n我完全明白这是为什么\nDeep Seek已拥有最强\n且最稳定的后训练方案之一\n批词不变性\n让强化学习流程的可预测性大幅提升\n此外还有多教师在线策略蒸流\n他们现在将其称为\n大规模同策略蒸流\nV4.1会向不同领域的\n40多个各有专攻的教师模型学习\n这也说明DeepSeek目前认为\n投入更多时间和资金优化数据\n比继续研发新的后训练算法\n更有价值\n尤其是本篇论文中\n他们十分重视\n强化学习数据的\n自动化合成与环境构建\n说到底\n我认为就是垃圾进\n垃圾出\n但这也意味着\n当前的后训练范式\n在一段时间内\n可能会少很多有意思的新进展\n不过客观来说\nMOPT也是最近才推出的\n所以这套范式问世的时间\n本来也不算长\n不过他们还做了一项\n我认为之前没有明确提到过的小调整\n那就是模型合并\n我也做过一期关于模型合并的视频\n在强化学习过程中\n他们会合并来自不同脚手架\n不同配置的独立训练运行的检查点\n整合不同优化路径的提升效果\n在用合并后的模型初始化\n下一次强化学习运行\n也就是说模型合并是用来持续汇总各次成功\n强化学习运行带来的提升等到流程最后\n他们会在不同领域和训练阶段开展大规模同策略蒸馏\n将所有专家教师模型的能力整合到同一个模型中\n另一方面推理力度也有一项升级\n它不再像V4那样只设置Think High和Think Max模式\n而是通过强化学习\n让推理力度成为一个从1到100的数值\n力度越低\nToken惩罚越强\n力度越高\n就会逐步允许模型使用更多算力\n因此\n模型不是训练几个彼此孤立的推理模式\n而是学会一条平滑的成本质量曲线\n面向用户的Low High和Max 3档\n基本上就是同一尺度上\n力度值为50,75和100的设置\n好啦\nV4.1 Flash在技术报告中提出的\n大部分重要进展就介绍到这里\n这份51页的技术报告里\n还有许多技术细节\n但其中有些内容实在太深奥\n或者需要更多背景知识才能解释清楚\n比如DSEC缩放多莫态混合专家均衡\n以及其他基础设施改进\n所以如果你们想了解更多\n一定要去看看他们的论文\n谢谢大家观看\n特别感谢Spamatch Chris Ledoux\nDegan\nRoberts Zaviasa\nPoof\nInu\nDX Research Group\nAlex\nLeo Halawani\n以及其他许多通过Patreon\n或YouTube支持我的人\n还没关注我的话到推特上关注我\n我们下期再见", "segments": [ { "start": 0.0, "end": 1.72, "text": "几个月前我们惊掉下巴", "raw_asr_text": "几个月前我们惊掉下巴", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 0, "words": [ { "word": "几", "start": 0.0, "end": 0.26, "probability": 0.8408203125 }, { "word": "个", "start": 0.26, "end": 0.36, "probability": 0.98681640625 }, { "word": "月", "start": 0.36, "end": 0.48, "probability": 0.9990234375 }, { "word": "前", "start": 0.48, "end": 0.72, "probability": 0.9951171875 }, { "word": "我们", "start": 0.72, "end": 0.98, "probability": 0.83642578125 }, { "word": "惊", "start": 0.98, "end": 1.18, "probability": 0.9794921875 }, { "word": "掉", "start": 1.18, "end": 1.34, "probability": 0.9873046875 }, { "word": "下", "start": 1.34, "end": 1.54, "probability": 0.99658203125 }, { "word": "巴", "start": 1.54, "end": 1.72, "probability": 0.98828125 } ], "asr_metrics": { "avg_logprob": -0.09112750723006878, "compression_ratio": 1.2357954545454546, "no_speech_prob": 1.9680021501322642e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "几个月前我们惊掉下巴当时发现Deep Seek基本上重新设计了整个V4模型架构核心就是让百万Token上下文便宜的离谱我们花了足足两期总共50分钟的系列视频拆解它们为打造这样一部杰作做了什么所以虽然我们本以为那已经是有史以来优化最激进的长上下文模型几周前我们却迎来了Deep Seek V4.1 Flash这款最新模型的架构改造甚至更疯狂V4.1" } ], "id": 0, "review_annotation": "", "row_id": 0, "bounded_row_id": 0, "parent_row_ids": [ 0 ], "parent_row_id": 0, "packet_id": 0, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 0, "packet_end": 28.38798557692308, "raw_text": "几个月前我们惊掉下巴", "selected_text": "几个月前我们惊掉下巴", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 0 ], "packet_id": 0, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1.72, "end": 3.14, "text": "当时发现Deep Seek", "raw_asr_text": "当时发现Deep Seek", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 0, "words": [ { "word": "当", "start": 1.72, "end": 1.96, "probability": 0.96240234375 }, { "word": "时", "start": 1.96, "end": 2.12, "probability": 0.9990234375 }, { "word": "发", "start": 2.12, "end": 2.3, "probability": 0.9970703125 }, { "word": "现", "start": 2.3, "end": 2.5, "probability": 0.99951171875 }, { "word": "De", "start": 2.5, "end": 2.72, "probability": 0.6875 }, { "word": "ep", "start": 2.72, "end": 2.82, "probability": 0.98291015625 }, { "word": " Se", "start": 2.82, "end": 3.0, "probability": 0.57470703125 }, { "word": "ek", "start": 3.0, "end": 3.14, "probability": 0.984375 } ], "asr_metrics": { "avg_logprob": -0.09112750723006878, "compression_ratio": 1.2357954545454546, "no_speech_prob": 1.9680021501322642e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "几个月前我们惊掉下巴当时发现Deep Seek基本上重新设计了整个V4模型架构核心就是让百万Token上下文便宜的离谱我们花了足足两期总共50分钟的系列视频拆解它们为打造这样一部杰作做了什么所以虽然我们本以为那已经是有史以来优化最激进的长上下文模型几周前我们却迎来了Deep Seek V4.1 Flash这款最新模型的架构改造甚至更疯狂V4.1" } ], "id": 1, "review_annotation": "", "row_id": 1, "bounded_row_id": 1, "parent_row_ids": [ 1 ], "parent_row_id": 1, "packet_id": 0, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 0, "packet_end": 28.38798557692308, "raw_text": "当时发现Deep Seek", "selected_text": "当时发现Deep Seek", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 1 ], "packet_id": 0, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 3.14, "end": 5.94, "text": "基本上重新设计了整个V4模型架构", "raw_asr_text": "基本上重新设计了整个V4模型架构", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 0, "words": [ { "word": "基本", "start": 3.14, "end": 3.66, "probability": 0.96923828125 }, { "word": "上", "start": 3.66, "end": 3.82, "probability": 0.99658203125 }, { "word": "重", "start": 3.82, "end": 4.0, "probability": 0.99267578125 }, { "word": "新", "start": 4.0, "end": 4.18, "probability": 0.99853515625 }, { "word": "设", "start": 4.18, "end": 4.34, "probability": 0.99853515625 }, { "word": "计", "start": 4.34, "end": 4.46, "probability": 0.9990234375 }, { "word": "了", "start": 4.46, "end": 4.58, "probability": 0.97900390625 }, { "word": "整", "start": 4.58, "end": 4.74, "probability": 0.99072265625 }, { "word": "个", "start": 4.74, "end": 4.94, "probability": 0.99853515625 }, { "word": "V", "start": 4.94, "end": 5.12, "probability": 0.90673828125 }, { "word": "4", "start": 5.12, "end": 5.3, "probability": 0.97412109375 }, { "word": "模", "start": 5.3, "end": 5.46, "probability": 0.986328125 }, { "word": "型", "start": 5.46, "end": 5.6, "probability": 0.99951171875 }, { "word": "架", "start": 5.6, "end": 5.76, "probability": 0.99755859375 }, { "word": "构", "start": 5.76, "end": 5.94, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.09112750723006878, "compression_ratio": 1.2357954545454546, "no_speech_prob": 1.9680021501322642e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "几个月前我们惊掉下巴当时发现Deep Seek基本上重新设计了整个V4模型架构核心就是让百万Token上下文便宜的离谱我们花了足足两期总共50分钟的系列视频拆解它们为打造这样一部杰作做了什么所以虽然我们本以为那已经是有史以来优化最激进的长上下文模型几周前我们却迎来了Deep Seek V4.1 Flash这款最新模型的架构改造甚至更疯狂V4.1" } ], "id": 2, "review_annotation": "", "row_id": 2, "bounded_row_id": 2, "parent_row_ids": [ 2 ], "parent_row_id": 2, "packet_id": 0, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 0, "packet_end": 28.38798557692308, "raw_text": "基本上重新设计了整个V4模型架构", "selected_text": "基本上重新设计了整个V4模型架构", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 2 ], "packet_id": 0, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 5.94, "end": 8.88, "text": "核心就是让百万Token上下文便宜的离谱", "raw_asr_text": "核心就是让百万Token上下文便宜的离谱", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 0, "words": [ { "word": "核", "start": 5.94, "end": 6.18, "probability": 0.986328125 }, { "word": "心", "start": 6.18, "end": 6.34, "probability": 0.99951171875 }, { "word": "就是", "start": 6.34, "end": 6.58, "probability": 0.9873046875 }, { "word": "让", "start": 6.58, "end": 6.8, "probability": 0.99072265625 }, { "word": "百", "start": 6.8, "end": 7.02, "probability": 0.9716796875 }, { "word": "万", "start": 7.02, "end": 7.24, "probability": 0.99951171875 }, { "word": "T", "start": 7.24, "end": 7.4, "probability": 0.71240234375 }, { "word": "oken", "start": 7.4, "end": 7.54, "probability": 0.97265625 }, { "word": "上", "start": 7.54, "end": 7.72, "probability": 0.9833984375 }, { "word": "下", "start": 7.72, "end": 7.86, "probability": 0.9990234375 }, { "word": "文", "start": 7.86, "end": 8.02, "probability": 0.87109375 }, { "word": "便", "start": 8.02, "end": 8.24, "probability": 0.9853515625 }, { "word": "宜", "start": 8.24, "end": 8.42, "probability": 0.99951171875 }, { "word": "的", "start": 8.42, "end": 8.52, "probability": 0.9375 }, { "word": "离", "start": 8.52, "end": 8.72, "probability": 0.99755859375 }, { "word": "谱", "start": 8.72, "end": 8.88, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.09112750723006878, "compression_ratio": 1.2357954545454546, "no_speech_prob": 1.9680021501322642e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "几个月前我们惊掉下巴当时发现Deep Seek基本上重新设计了整个V4模型架构核心就是让百万Token上下文便宜的离谱我们花了足足两期总共50分钟的系列视频拆解它们为打造这样一部杰作做了什么所以虽然我们本以为那已经是有史以来优化最激进的长上下文模型几周前我们却迎来了Deep Seek V4.1 Flash这款最新模型的架构改造甚至更疯狂V4.1" } ], "id": 3, "review_annotation": "", "row_id": 3, "bounded_row_id": 3, "parent_row_ids": [ 3 ], "parent_row_id": 3, "packet_id": 0, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 0, "packet_end": 28.38798557692308, "raw_text": "核心就是让百万Token上下文便宜的离谱", "selected_text": "核心就是让百万Token上下文便宜的离谱", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 3 ], "packet_id": 0, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 8.88, "end": 13.0, "text": "我们花了足足两期总共50分钟的系列视频", "raw_asr_text": "我们花了足足两期总共50分钟的系列视频", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 0, "words": [ { "word": "我们", "start": 8.88, "end": 9.36, "probability": 0.99560546875 }, { "word": "花", "start": 9.36, "end": 9.7, "probability": 0.99169921875 }, { "word": "了", "start": 9.7, "end": 9.88, "probability": 0.9970703125 }, { "word": "足", "start": 9.88, "end": 10.08, "probability": 0.98388671875 }, { "word": "足", "start": 10.08, "end": 10.32, "probability": 0.998046875 }, { "word": "两", "start": 10.32, "end": 10.58, "probability": 0.9580078125 }, { "word": "期", "start": 10.58, "end": 10.92, "probability": 0.9990234375 }, { "word": "总", "start": 10.92, "end": 11.14, "probability": 0.95703125 }, { "word": "共", "start": 11.14, "end": 11.34, "probability": 0.99853515625 }, { "word": "50", "start": 11.34, "end": 11.74, "probability": 0.92919921875 }, { "word": "分", "start": 11.74, "end": 11.92, "probability": 0.99365234375 }, { "word": "钟", "start": 11.92, "end": 12.14, "probability": 0.99755859375 }, { "word": "的", "start": 12.14, "end": 12.28, "probability": 0.99462890625 }, { "word": "系", "start": 12.28, "end": 12.48, "probability": 0.97021484375 }, { "word": "列", "start": 12.48, "end": 12.62, "probability": 0.99951171875 }, { "word": "视", "start": 12.62, "end": 12.8, "probability": 0.9990234375 }, { "word": "频", "start": 12.8, "end": 13.0, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09112750723006878, "compression_ratio": 1.2357954545454546, "no_speech_prob": 1.9680021501322642e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "几个月前我们惊掉下巴当时发现Deep Seek基本上重新设计了整个V4模型架构核心就是让百万Token上下文便宜的离谱我们花了足足两期总共50分钟的系列视频拆解它们为打造这样一部杰作做了什么所以虽然我们本以为那已经是有史以来优化最激进的长上下文模型几周前我们却迎来了Deep Seek V4.1 Flash这款最新模型的架构改造甚至更疯狂V4.1" } ], "id": 4, "review_annotation": "", "row_id": 4, "bounded_row_id": 4, "parent_row_ids": [ 4 ], "parent_row_id": 4, "packet_id": 0, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 0, "packet_end": 28.38798557692308, "raw_text": "我们花了足足两期总共50分钟的系列视频", "selected_text": "我们花了足足两期总共50分钟的系列视频", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 4 ], "packet_id": 0, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 13.0, "end": 15.22, "text": "拆解它们为打造这样一部杰作", "raw_asr_text": "拆解它们为打造这样一部杰作", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 0, "words": [ { "word": "拆", "start": 13.0, "end": 13.4, "probability": 0.99609375 }, { "word": "解", "start": 13.4, "end": 13.58, "probability": 0.9990234375 }, { "word": "它", "start": 13.58, "end": 13.72, "probability": 0.5673828125 }, { "word": "们", "start": 13.72, "end": 13.84, "probability": 0.99267578125 }, { "word": "为", "start": 13.84, "end": 13.98, "probability": 0.994140625 }, { "word": "打", "start": 13.98, "end": 14.16, "probability": 0.99755859375 }, { "word": "造", "start": 14.16, "end": 14.34, "probability": 0.99951171875 }, { "word": "这样", "start": 14.34, "end": 14.56, "probability": 0.9814453125 }, { "word": "一", "start": 14.56, "end": 14.74, "probability": 0.9951171875 }, { "word": "部", "start": 14.74, "end": 14.84, "probability": 0.9814453125 }, { "word": "杰", "start": 14.84, "end": 15.06, "probability": 0.89453125 }, { "word": "作", "start": 15.06, "end": 15.22, "probability": 0.9892578125 } ], "asr_metrics": { "avg_logprob": -0.09112750723006878, "compression_ratio": 1.2357954545454546, "no_speech_prob": 1.9680021501322642e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "几个月前我们惊掉下巴当时发现Deep Seek基本上重新设计了整个V4模型架构核心就是让百万Token上下文便宜的离谱我们花了足足两期总共50分钟的系列视频拆解它们为打造这样一部杰作做了什么所以虽然我们本以为那已经是有史以来优化最激进的长上下文模型几周前我们却迎来了Deep Seek V4.1 Flash这款最新模型的架构改造甚至更疯狂V4.1" } ], "id": 5, "review_annotation": "", "row_id": 5, "bounded_row_id": 5, "parent_row_ids": [ 5 ], "parent_row_id": 5, "packet_id": 0, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 0, "packet_end": 28.38798557692308, "raw_text": "拆解它们为打造这样一部杰作", "selected_text": "拆解它们为打造这样一部杰作", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 5 ], "packet_id": 0, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 15.22, "end": 15.94, "text": "做了什么", "raw_asr_text": "做了什么", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 0, "words": [ { "word": "做", "start": 15.22, "end": 15.58, "probability": 0.9482421875 }, { "word": "了", "start": 15.58, "end": 15.72, "probability": 0.99951171875 }, { "word": "什么", "start": 15.72, "end": 15.94, "probability": 0.99560546875 } ], "asr_metrics": { "avg_logprob": -0.09112750723006878, "compression_ratio": 1.2357954545454546, "no_speech_prob": 1.9680021501322642e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "几个月前我们惊掉下巴当时发现Deep Seek基本上重新设计了整个V4模型架构核心就是让百万Token上下文便宜的离谱我们花了足足两期总共50分钟的系列视频拆解它们为打造这样一部杰作做了什么所以虽然我们本以为那已经是有史以来优化最激进的长上下文模型几周前我们却迎来了Deep Seek V4.1 Flash这款最新模型的架构改造甚至更疯狂V4.1" } ], "id": 6, "review_annotation": "", "row_id": 6, "bounded_row_id": 6, "parent_row_ids": [ 6 ], "parent_row_id": 6, "packet_id": 0, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 0, "packet_end": 28.38798557692308, "raw_text": "做了什么", "selected_text": "做了什么", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 6 ], "packet_id": 0, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 15.94, "end": 18.02, "text": "所以虽然我们本以为那已经是", "raw_asr_text": "所以虽然我们本以为那已经是", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 0, "words": [ { "word": "所以", "start": 15.94, "end": 16.38, "probability": 0.8876953125 }, { "word": "虽", "start": 16.38, "end": 16.68, "probability": 0.9970703125 }, { "word": "然", "start": 16.68, "end": 16.82, "probability": 0.99951171875 }, { "word": "我们", "start": 16.82, "end": 17.0, "probability": 0.99462890625 }, { "word": "本", "start": 17.0, "end": 17.24, "probability": 0.99853515625 }, { "word": "以", "start": 17.24, "end": 17.4, "probability": 0.9970703125 }, { "word": "为", "start": 17.4, "end": 17.5, "probability": 0.9990234375 }, { "word": "那", "start": 17.5, "end": 17.66, "probability": 0.9833984375 }, { "word": "已经", "start": 17.66, "end": 17.86, "probability": 0.9970703125 }, { "word": "是", "start": 17.86, "end": 18.02, "probability": 0.9931640625 } ], "asr_metrics": { "avg_logprob": -0.09112750723006878, "compression_ratio": 1.2357954545454546, "no_speech_prob": 1.9680021501322642e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "几个月前我们惊掉下巴当时发现Deep Seek基本上重新设计了整个V4模型架构核心就是让百万Token上下文便宜的离谱我们花了足足两期总共50分钟的系列视频拆解它们为打造这样一部杰作做了什么所以虽然我们本以为那已经是有史以来优化最激进的长上下文模型几周前我们却迎来了Deep Seek V4.1 Flash这款最新模型的架构改造甚至更疯狂V4.1" } ], "id": 7, "review_annotation": "", "row_id": 7, "bounded_row_id": 7, "parent_row_ids": [ 7 ], "parent_row_id": 7, "packet_id": 0, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 0, "packet_end": 28.38798557692308, "raw_text": "所以虽然我们本以为那已经是", "selected_text": "所以虽然我们本以为那已经是", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 7 ], "packet_id": 0, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 18.02, "end": 20.82, "text": "有史以来优化最激进的长上下文模型", "raw_asr_text": "有史以来优化最激进的长上下文模型", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 0, "words": [ { "word": "有", "start": 18.02, "end": 18.24, "probability": 0.98876953125 }, { "word": "史", "start": 18.24, "end": 18.36, "probability": 0.91064453125 }, { "word": "以", "start": 18.36, "end": 18.5, "probability": 0.9970703125 }, { "word": "来", "start": 18.5, "end": 18.6, "probability": 0.99951171875 }, { "word": "优", "start": 18.6, "end": 18.74, "probability": 0.9990234375 }, { "word": "化", "start": 18.74, "end": 18.88, "probability": 0.99951171875 }, { "word": "最", "start": 18.88, "end": 19.12, "probability": 0.9951171875 }, { "word": "激", "start": 19.12, "end": 19.36, "probability": 0.9892578125 }, { "word": "进", "start": 19.36, "end": 19.52, "probability": 0.99658203125 }, { "word": "的", "start": 19.52, "end": 19.64, "probability": 0.99853515625 }, { "word": "长", "start": 19.64, "end": 19.88, "probability": 0.84423828125 }, { "word": "上", "start": 19.88, "end": 20.1, "probability": 0.99267578125 }, { "word": "下", "start": 20.1, "end": 20.26, "probability": 0.9990234375 }, { "word": "文", "start": 20.26, "end": 20.42, "probability": 0.99462890625 }, { "word": "模", "start": 20.42, "end": 20.58, "probability": 0.98779296875 }, { "word": "型", "start": 20.58, "end": 20.82, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.09112750723006878, "compression_ratio": 1.2357954545454546, "no_speech_prob": 1.9680021501322642e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "几个月前我们惊掉下巴当时发现Deep Seek基本上重新设计了整个V4模型架构核心就是让百万Token上下文便宜的离谱我们花了足足两期总共50分钟的系列视频拆解它们为打造这样一部杰作做了什么所以虽然我们本以为那已经是有史以来优化最激进的长上下文模型几周前我们却迎来了Deep Seek V4.1 Flash这款最新模型的架构改造甚至更疯狂V4.1" } ], "id": 8, "review_annotation": "", "row_id": 8, "bounded_row_id": 8, "parent_row_ids": [ 8 ], "parent_row_id": 8, "packet_id": 0, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 0, "packet_end": 28.38798557692308, "raw_text": "有史以来优化最激进的长上下文模型", "selected_text": "有史以来优化最激进的长上下文模型", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 8 ], "packet_id": 0, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 20.82, "end": 24.28, "text": "几周前我们却迎来了Deep Seek V4.1 Flash", "raw_asr_text": "几周前我们却迎来了Deep Seek V4.1 Flash", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 0, "words": [ { "word": "几", "start": 20.82, "end": 21.18, "probability": 0.998046875 }, { "word": "周", "start": 21.18, "end": 21.3, "probability": 0.99853515625 }, { "word": "前", "start": 21.3, "end": 21.48, "probability": 0.99462890625 }, { "word": "我们", "start": 21.48, "end": 21.66, "probability": 0.99658203125 }, { "word": "却", "start": 21.66, "end": 21.84, "probability": 0.99072265625 }, { "word": "迎", "start": 21.84, "end": 21.96, "probability": 0.998046875 }, { "word": "来了", "start": 21.96, "end": 22.18, "probability": 0.98583984375 }, { "word": "De", "start": 22.18, "end": 22.38, "probability": 0.982421875 }, { "word": "ep", "start": 22.38, "end": 22.5, "probability": 0.9990234375 }, { "word": " Se", "start": 22.5, "end": 22.66, "probability": 0.98486328125 }, { "word": "ek", "start": 22.66, "end": 22.82, "probability": 1.0 }, { "word": " V", "start": 22.82, "end": 23.06, "probability": 0.7724609375 }, { "word": "4.", "start": 23.06, "end": 23.42, "probability": 0.994140625 }, { "word": "1", "start": 23.62, "end": 23.9, "probability": 0.99267578125 }, { "word": " Flash", "start": 23.9, "end": 24.28, "probability": 0.85400390625 } ], "asr_metrics": { "avg_logprob": -0.09112750723006878, "compression_ratio": 1.2357954545454546, "no_speech_prob": 1.9680021501322642e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "几个月前我们惊掉下巴当时发现Deep Seek基本上重新设计了整个V4模型架构核心就是让百万Token上下文便宜的离谱我们花了足足两期总共50分钟的系列视频拆解它们为打造这样一部杰作做了什么所以虽然我们本以为那已经是有史以来优化最激进的长上下文模型几周前我们却迎来了Deep Seek V4.1 Flash这款最新模型的架构改造甚至更疯狂V4.1" } ], "id": 9, "review_annotation": "", "row_id": 9, "bounded_row_id": 9, "parent_row_ids": [ 9 ], "parent_row_id": 9, "packet_id": 0, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 0, "packet_end": 28.38798557692308, "raw_text": "几周前我们却迎来了Deep Seek V4.1 Flash", "selected_text": "几周前我们却迎来了Deep Seek V4.1 Flash", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 9 ], "packet_id": 0, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 24.7, "end": 27.587985576923078, "text": "这款最新模型的架构改造甚至更疯狂", "raw_asr_text": "这款最新模型的架构改造甚至更疯狂", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 0, "words": [ { "word": "这", "start": 24.7, "end": 25.06, "probability": 0.9931640625 }, { "word": "款", "start": 25.06, "end": 25.22, "probability": 0.9345703125 }, { "word": "最", "start": 25.22, "end": 25.36, "probability": 0.99560546875 }, { "word": "新", "start": 25.36, "end": 25.5, "probability": 0.99951171875 }, { "word": "模", "start": 25.5, "end": 25.62, "probability": 0.9755859375 }, { "word": "型", "start": 25.62, "end": 25.74, "probability": 0.99951171875 }, { "word": "的", "start": 25.74, "end": 25.82, "probability": 0.9892578125 }, { "word": "架", "start": 25.82, "end": 25.98, "probability": 0.99609375 }, { "word": "构", "start": 25.98, "end": 26.1, "probability": 0.9990234375 }, { "word": "改", "start": 26.1, "end": 26.2, "probability": 0.984375 }, { "word": "造", "start": 26.2, "end": 26.38, "probability": 0.99951171875 }, { "word": "甚", "start": 26.38, "end": 26.62, "probability": 0.9921875 }, { "word": "至", "start": 26.62, "end": 26.86, "probability": 1.0 }, { "word": "更", "start": 26.86, "end": 27.1, "probability": 0.99658203125 }, { "word": "疯", "start": 27.1, "end": 27.34, "probability": 0.998046875 }, { "word": "狂", "start": 27.34, "end": 27.6, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09112750723006878, "compression_ratio": 1.2357954545454546, "no_speech_prob": 1.9680021501322642e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "几个月前我们惊掉下巴当时发现Deep Seek基本上重新设计了整个V4模型架构核心就是让百万Token上下文便宜的离谱我们花了足足两期总共50分钟的系列视频拆解它们为打造这样一部杰作做了什么所以虽然我们本以为那已经是有史以来优化最激进的长上下文模型几周前我们却迎来了Deep Seek V4.1 Flash这款最新模型的架构改造甚至更疯狂V4.1" } ], "id": 10, "review_annotation": "", "row_id": 10, "bounded_row_id": 10, "parent_row_ids": [ 10 ], "parent_row_id": 10, "packet_id": 0, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 0, "packet_end": 28.38798557692308, "raw_text": "这款最新模型的架构改造甚至更疯狂", "selected_text": "这款最新模型的架构改造甚至更疯狂", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 10 ], "packet_id": 0, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 27.607985576923078, "end": 29.667985576923076, "text": "V4.1Flash的参数量", "raw_asr_text": "V4.1Flash的参数量", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 1, "words": [ { "word": "V", "start": 27.607985576923078, "end": 28.087985576923078, "probability": 0.37646484375 }, { "word": "4.", "start": 28.087985576923078, "end": 28.287985576923077, "probability": 0.98193359375 }, { "word": "1", "start": 28.427985576923078, "end": 28.587985576923078, "probability": 0.97900390625 }, { "word": "Fl", "start": 28.587985576923078, "end": 28.767985576923078, "probability": 0.4189453125 }, { "word": "ash", "start": 28.767985576923078, "end": 28.927985576923078, "probability": 0.9033203125 }, { "word": "的", "start": 28.927985576923078, "end": 29.06798557692308, "probability": 0.9755859375 }, { "word": "参", "start": 29.06798557692308, "end": 29.327985576923076, "probability": 0.95166015625 }, { "word": "数", "start": 29.327985576923076, "end": 29.487985576923077, "probability": 0.99853515625 }, { "word": "量", "start": 29.487985576923077, "end": 29.667985576923076, "probability": 0.9951171875 } ], "asr_metrics": { "avg_logprob": -0.16982789856294242, "compression_ratio": 1.1712538226299694, "no_speech_prob": 5.825580295937272e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "更疯狂V4.1Flash的参数量几乎是V4Flash的2倍从2840亿增长到5520亿而且其新架构还做了多方面精简足以把KV缓存再压缩4倍为了让你明白这有多厉害过去9个月里DeepSeek已经把每个Token的KV缓存大小压缩了54倍实在太惊人了所以在今天的视频里我们就来看看DeepSeek是如何降低前沿AI模型成本的以及他们想" } ], "id": 11, "review_annotation": "", "row_id": 11, "bounded_row_id": 11, "parent_row_ids": [ 11 ], "parent_row_id": 11, "packet_id": 1, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 26.787985576923077, "packet_end": 55.97597115384615, "raw_text": "V4.1Flash的参数量", "selected_text": "V4.1Flash的参数量", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 11 ], "packet_id": 1, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 29.667985576923076, "end": 31.627985576923077, "text": "几乎是V4Flash的2倍", "raw_asr_text": "几乎是V4Flash的2倍", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 1, "words": [ { "word": "几", "start": 29.667985576923076, "end": 29.987985576923077, "probability": 0.97265625 }, { "word": "乎", "start": 29.987985576923077, "end": 30.107985576923078, "probability": 1.0 }, { "word": "是", "start": 30.107985576923078, "end": 30.247985576923078, "probability": 0.98681640625 }, { "word": "V", "start": 30.247985576923078, "end": 30.447985576923077, "probability": 0.955078125 }, { "word": "4", "start": 30.447985576923077, "end": 30.647985576923077, "probability": 0.9931640625 }, { "word": "Fl", "start": 30.647985576923077, "end": 30.827985576923076, "probability": 0.873046875 }, { "word": "ash", "start": 30.827985576923076, "end": 31.007985576923076, "probability": 0.9990234375 }, { "word": "的", "start": 31.007985576923076, "end": 31.147985576923077, "probability": 0.9775390625 }, { "word": "2", "start": 31.147985576923077, "end": 31.347985576923076, "probability": 0.483642578125 }, { "word": "倍", "start": 31.347985576923076, "end": 31.627985576923077, "probability": 0.9970703125 } ], "asr_metrics": { "avg_logprob": -0.16982789856294242, "compression_ratio": 1.1712538226299694, "no_speech_prob": 5.825580295937272e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "更疯狂V4.1Flash的参数量几乎是V4Flash的2倍从2840亿增长到5520亿而且其新架构还做了多方面精简足以把KV缓存再压缩4倍为了让你明白这有多厉害过去9个月里DeepSeek已经把每个Token的KV缓存大小压缩了54倍实在太惊人了所以在今天的视频里我们就来看看DeepSeek是如何降低前沿AI模型成本的以及他们想" } ], "id": 12, "review_annotation": "", "row_id": 12, "bounded_row_id": 12, "parent_row_ids": [ 12 ], "parent_row_id": 12, "packet_id": 1, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 26.787985576923077, "packet_end": 55.97597115384615, "raw_text": "几乎是V4Flash的2倍", "selected_text": "几乎是V4Flash的2倍", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 12 ], "packet_id": 1, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 31.627985576923077, "end": 35.12798557692308, "text": "从2840亿增长到5520亿", "raw_asr_text": "从2840亿增长到5520亿", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 1, "words": [ { "word": "从", "start": 31.627985576923077, "end": 32.12798557692308, "probability": 0.87353515625 }, { "word": "28", "start": 32.12798557692308, "end": 32.38798557692308, "probability": 0.9658203125 }, { "word": "40", "start": 32.38798557692308, "end": 33.187985576923076, "probability": 0.9931640625 }, { "word": "亿", "start": 33.187985576923076, "end": 33.327985576923076, "probability": 0.990234375 }, { "word": "增", "start": 33.327985576923076, "end": 33.52798557692308, "probability": 0.9951171875 }, { "word": "长", "start": 33.52798557692308, "end": 33.687985576923076, "probability": 0.9970703125 }, { "word": "到", "start": 33.687985576923076, "end": 33.827985576923076, "probability": 0.9775390625 }, { "word": "55", "start": 33.827985576923076, "end": 34.227985576923075, "probability": 0.97314453125 }, { "word": "20", "start": 34.227985576923075, "end": 34.84798557692308, "probability": 0.998046875 }, { "word": "亿", "start": 34.84798557692308, "end": 35.12798557692308, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.16982789856294242, "compression_ratio": 1.1712538226299694, "no_speech_prob": 5.825580295937272e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "更疯狂V4.1Flash的参数量几乎是V4Flash的2倍从2840亿增长到5520亿而且其新架构还做了多方面精简足以把KV缓存再压缩4倍为了让你明白这有多厉害过去9个月里DeepSeek已经把每个Token的KV缓存大小压缩了54倍实在太惊人了所以在今天的视频里我们就来看看DeepSeek是如何降低前沿AI模型成本的以及他们想" } ], "id": 13, "review_annotation": "", "row_id": 13, "bounded_row_id": 13, "parent_row_ids": [ 13 ], "parent_row_id": 13, "packet_id": 1, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 26.787985576923077, "packet_end": 55.97597115384615, "raw_text": "从2840亿增长到5520亿", "selected_text": "从2840亿增长到5520亿", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 13 ], "packet_id": 1, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 35.12798557692308, "end": 37.727985576923075, "text": "而且其新架构还做了多方面精简", "raw_asr_text": "而且其新架构还做了多方面精简", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 1, "words": [ { "word": "而且", "start": 35.12798557692308, "end": 35.52798557692308, "probability": 0.99169921875 }, { "word": "其", "start": 35.52798557692308, "end": 35.727985576923075, "probability": 0.978515625 }, { "word": "新", "start": 35.727985576923075, "end": 35.96798557692308, "probability": 0.92919921875 }, { "word": "架", "start": 35.96798557692308, "end": 36.16798557692308, "probability": 0.9853515625 }, { "word": "构", "start": 36.16798557692308, "end": 36.34798557692308, "probability": 0.9951171875 }, { "word": "还", "start": 36.34798557692308, "end": 36.48798557692308, "probability": 0.96875 }, { "word": "做", "start": 36.48798557692308, "end": 36.64798557692308, "probability": 0.986328125 }, { "word": "了", "start": 36.64798557692308, "end": 36.78798557692308, "probability": 0.9912109375 }, { "word": "多", "start": 36.78798557692308, "end": 36.96798557692308, "probability": 0.99072265625 }, { "word": "方面", "start": 36.96798557692308, "end": 37.327985576923076, "probability": 0.990234375 }, { "word": "精", "start": 37.327985576923076, "end": 37.52798557692308, "probability": 0.6669921875 }, { "word": "简", "start": 37.52798557692308, "end": 37.727985576923075, "probability": 0.7998046875 } ], "asr_metrics": { "avg_logprob": -0.16982789856294242, "compression_ratio": 1.1712538226299694, "no_speech_prob": 5.825580295937272e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "更疯狂V4.1Flash的参数量几乎是V4Flash的2倍从2840亿增长到5520亿而且其新架构还做了多方面精简足以把KV缓存再压缩4倍为了让你明白这有多厉害过去9个月里DeepSeek已经把每个Token的KV缓存大小压缩了54倍实在太惊人了所以在今天的视频里我们就来看看DeepSeek是如何降低前沿AI模型成本的以及他们想" } ], "id": 14, "review_annotation": "", "row_id": 14, "bounded_row_id": 14, "parent_row_ids": [ 14 ], "parent_row_id": 14, "packet_id": 1, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 26.787985576923077, "packet_end": 55.97597115384615, "raw_text": "而且其新架构还做了多方面精简", "selected_text": "而且其新架构还做了多方面精简", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 14 ], "packet_id": 1, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 37.727985576923075, "end": 40.02798557692308, "text": "足以把KV缓存再压缩4倍", "raw_asr_text": "足以把KV缓存再压缩4倍", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 1, "words": [ { "word": "足", "start": 37.727985576923075, "end": 37.947985576923074, "probability": 0.99365234375 }, { "word": "以", "start": 37.947985576923074, "end": 38.047985576923075, "probability": 0.9912109375 }, { "word": "把", "start": 38.047985576923075, "end": 38.20798557692308, "probability": 0.99072265625 }, { "word": "K", "start": 38.20798557692308, "end": 38.407985576923075, "probability": 0.87890625 }, { "word": "V", "start": 38.407985576923075, "end": 38.56798557692308, "probability": 0.9912109375 }, { "word": "缓", "start": 38.56798557692308, "end": 38.74798557692308, "probability": 0.99755859375 }, { "word": "存", "start": 38.74798557692308, "end": 38.96798557692308, "probability": 0.99951171875 }, { "word": "再", "start": 38.96798557692308, "end": 39.30798557692307, "probability": 0.91796875 }, { "word": "压", "start": 39.30798557692307, "end": 39.547985576923075, "probability": 0.99951171875 }, { "word": "缩", "start": 39.547985576923075, "end": 39.687985576923076, "probability": 0.99853515625 }, { "word": "4", "start": 39.687985576923076, "end": 39.88798557692308, "probability": 0.90625 }, { "word": "倍", "start": 39.88798557692308, "end": 40.02798557692308, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.16982789856294242, "compression_ratio": 1.1712538226299694, "no_speech_prob": 5.825580295937272e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "更疯狂V4.1Flash的参数量几乎是V4Flash的2倍从2840亿增长到5520亿而且其新架构还做了多方面精简足以把KV缓存再压缩4倍为了让你明白这有多厉害过去9个月里DeepSeek已经把每个Token的KV缓存大小压缩了54倍实在太惊人了所以在今天的视频里我们就来看看DeepSeek是如何降低前沿AI模型成本的以及他们想" } ], "id": 15, "review_annotation": "", "row_id": 15, "bounded_row_id": 15, "parent_row_ids": [ 15 ], "parent_row_id": 15, "packet_id": 1, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 26.787985576923077, "packet_end": 55.97597115384615, "raw_text": "足以把KV缓存再压缩4倍", "selected_text": "足以把KV缓存再压缩4倍", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 15 ], "packet_id": 1, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 40.02798557692308, "end": 42.007985576923076, "text": "为了让你明白这有多厉害", "raw_asr_text": "为了让你明白这有多厉害", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 1, "words": [ { "word": "为", "start": 40.02798557692308, "end": 40.327985576923076, "probability": 0.978515625 }, { "word": "了", "start": 40.327985576923076, "end": 40.447985576923074, "probability": 0.99951171875 }, { "word": "让", "start": 40.447985576923074, "end": 40.56798557692308, "probability": 0.99560546875 }, { "word": "你", "start": 40.56798557692308, "end": 40.70798557692308, "probability": 0.97705078125 }, { "word": "明白", "start": 40.70798557692308, "end": 41.047985576923075, "probability": 0.9892578125 }, { "word": "这", "start": 41.047985576923075, "end": 41.24798557692308, "probability": 0.974609375 }, { "word": "有", "start": 41.24798557692308, "end": 41.34798557692308, "probability": 0.99560546875 }, { "word": "多", "start": 41.34798557692308, "end": 41.52798557692308, "probability": 0.9990234375 }, { "word": "厉", "start": 41.52798557692308, "end": 41.80798557692307, "probability": 0.9990234375 }, { "word": "害", "start": 41.80798557692307, "end": 42.007985576923076, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.16982789856294242, "compression_ratio": 1.1712538226299694, "no_speech_prob": 5.825580295937272e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "更疯狂V4.1Flash的参数量几乎是V4Flash的2倍从2840亿增长到5520亿而且其新架构还做了多方面精简足以把KV缓存再压缩4倍为了让你明白这有多厉害过去9个月里DeepSeek已经把每个Token的KV缓存大小压缩了54倍实在太惊人了所以在今天的视频里我们就来看看DeepSeek是如何降低前沿AI模型成本的以及他们想" } ], "id": 16, "review_annotation": "", "row_id": 16, "bounded_row_id": 16, "parent_row_ids": [ 16 ], "parent_row_id": 16, "packet_id": 1, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 26.787985576923077, "packet_end": 55.97597115384615, "raw_text": "为了让你明白这有多厉害", "selected_text": "为了让你明白这有多厉害", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 16 ], "packet_id": 1, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 42.007985576923076, "end": 43.06798557692308, "text": "过去9个月里", "raw_asr_text": "过去9个月里", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 1, "words": [ { "word": "过", "start": 42.007985576923076, "end": 42.30798557692307, "probability": 0.91455078125 }, { "word": "去", "start": 42.30798557692307, "end": 42.447985576923074, "probability": 0.99951171875 }, { "word": "9", "start": 42.447985576923074, "end": 42.60798557692308, "probability": 0.85009765625 }, { "word": "个", "start": 42.60798557692308, "end": 42.727985576923075, "probability": 0.998046875 }, { "word": "月", "start": 42.727985576923075, "end": 42.90798557692308, "probability": 0.99951171875 }, { "word": "里", "start": 42.90798557692308, "end": 43.06798557692308, "probability": 0.98828125 } ], "asr_metrics": { "avg_logprob": -0.16982789856294242, "compression_ratio": 1.1712538226299694, "no_speech_prob": 5.825580295937272e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "更疯狂V4.1Flash的参数量几乎是V4Flash的2倍从2840亿增长到5520亿而且其新架构还做了多方面精简足以把KV缓存再压缩4倍为了让你明白这有多厉害过去9个月里DeepSeek已经把每个Token的KV缓存大小压缩了54倍实在太惊人了所以在今天的视频里我们就来看看DeepSeek是如何降低前沿AI模型成本的以及他们想" } ], "id": 17, "review_annotation": "", "row_id": 17, "bounded_row_id": 17, "parent_row_ids": [ 17 ], "parent_row_id": 17, "packet_id": 1, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 26.787985576923077, "packet_end": 55.97597115384615, "raw_text": "过去9个月里", "selected_text": "过去9个月里", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 17 ], "packet_id": 1, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 43.06798557692308, "end": 45.42798557692308, "text": "DeepSeek已经把每个Token的KV", "raw_asr_text": "DeepSeek已经把每个Token的KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 1, "words": [ { "word": "De", "start": 43.06798557692308, "end": 43.30798557692307, "probability": 0.8232421875 }, { "word": "ep", "start": 43.30798557692307, "end": 43.38798557692308, "probability": 0.9296875 }, { "word": "Se", "start": 43.38798557692308, "end": 43.52798557692307, "probability": 0.3740234375 }, { "word": "ek", "start": 43.52798557692307, "end": 43.64798557692308, "probability": 0.6650390625 }, { "word": "已经", "start": 43.64798557692308, "end": 43.98798557692308, "probability": 0.98779296875 }, { "word": "把", "start": 43.98798557692308, "end": 44.227985576923075, "probability": 0.9921875 }, { "word": "每", "start": 44.227985576923075, "end": 44.46798557692308, "probability": 0.99755859375 }, { "word": "个", "start": 44.46798557692308, "end": 44.64798557692308, "probability": 0.990234375 }, { "word": "T", "start": 44.64798557692308, "end": 44.78798557692308, "probability": 0.859375 }, { "word": "oken", "start": 44.78798557692308, "end": 44.947985576923074, "probability": 0.97314453125 }, { "word": "的", "start": 44.947985576923074, "end": 45.06798557692308, "probability": 0.94970703125 }, { "word": "K", "start": 45.06798557692308, "end": 45.227985576923075, "probability": 0.9892578125 }, { "word": "V", "start": 45.227985576923075, "end": 45.42798557692308, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.16982789856294242, "compression_ratio": 1.1712538226299694, "no_speech_prob": 5.825580295937272e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "更疯狂V4.1Flash的参数量几乎是V4Flash的2倍从2840亿增长到5520亿而且其新架构还做了多方面精简足以把KV缓存再压缩4倍为了让你明白这有多厉害过去9个月里DeepSeek已经把每个Token的KV缓存大小压缩了54倍实在太惊人了所以在今天的视频里我们就来看看DeepSeek是如何降低前沿AI模型成本的以及他们想" } ], "id": 18, "review_annotation": "", "row_id": 18, "bounded_row_id": 18, "parent_row_ids": [ 18 ], "parent_row_id": 18, "packet_id": 1, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 26.787985576923077, "packet_end": 55.97597115384615, "raw_text": "DeepSeek已经把每个Token的KV", "selected_text": "DeepSeek已经把每个Token的KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 18 ], "packet_id": 1, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 45.42798557692308, "end": 47.84798557692308, "text": "缓存大小压缩了54倍", "raw_asr_text": "缓存大小压缩了54倍", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 1, "words": [ { "word": "缓", "start": 45.42798557692308, "end": 45.727985576923075, "probability": 0.99462890625 }, { "word": "存", "start": 45.727985576923075, "end": 45.90798557692308, "probability": 0.9990234375 }, { "word": "大", "start": 45.90798557692308, "end": 46.06798557692308, "probability": 0.99560546875 }, { "word": "小", "start": 46.06798557692308, "end": 46.30798557692307, "probability": 0.99853515625 }, { "word": "压", "start": 46.30798557692307, "end": 46.587985576923074, "probability": 0.998046875 }, { "word": "缩", "start": 46.587985576923074, "end": 46.78798557692308, "probability": 0.99951171875 }, { "word": "了", "start": 46.78798557692308, "end": 46.88798557692308, "probability": 0.96337890625 }, { "word": "54", "start": 46.88798557692308, "end": 47.42798557692308, "probability": 0.9912109375 }, { "word": "倍", "start": 47.42798557692308, "end": 47.84798557692308, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.16982789856294242, "compression_ratio": 1.1712538226299694, "no_speech_prob": 5.825580295937272e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "更疯狂V4.1Flash的参数量几乎是V4Flash的2倍从2840亿增长到5520亿而且其新架构还做了多方面精简足以把KV缓存再压缩4倍为了让你明白这有多厉害过去9个月里DeepSeek已经把每个Token的KV缓存大小压缩了54倍实在太惊人了所以在今天的视频里我们就来看看DeepSeek是如何降低前沿AI模型成本的以及他们想" } ], "id": 19, "review_annotation": "", "row_id": 19, "bounded_row_id": 19, "parent_row_ids": [ 19 ], "parent_row_id": 19, "packet_id": 1, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 26.787985576923077, "packet_end": 55.97597115384615, "raw_text": "缓存大小压缩了54倍", "selected_text": "缓存大小压缩了54倍", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 19 ], "packet_id": 1, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 47.84798557692308, "end": 49.70798557692308, "text": "实在太惊人了", "raw_asr_text": "实在太惊人了", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 1, "words": [ { "word": "实", "start": 47.84798557692308, "end": 48.447985576923074, "probability": 0.93798828125 }, { "word": "在", "start": 48.447985576923074, "end": 48.64798557692308, "probability": 0.99951171875 }, { "word": "太", "start": 48.64798557692308, "end": 48.947985576923074, "probability": 0.9921875 }, { "word": "惊", "start": 48.947985576923074, "end": 49.24798557692308, "probability": 0.9970703125 }, { "word": "人", "start": 49.24798557692308, "end": 49.46798557692308, "probability": 0.99951171875 }, { "word": "了", "start": 49.46798557692308, "end": 49.70798557692308, "probability": 0.9951171875 } ], "asr_metrics": { "avg_logprob": -0.16982789856294242, "compression_ratio": 1.1712538226299694, "no_speech_prob": 5.825580295937272e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "更疯狂V4.1Flash的参数量几乎是V4Flash的2倍从2840亿增长到5520亿而且其新架构还做了多方面精简足以把KV缓存再压缩4倍为了让你明白这有多厉害过去9个月里DeepSeek已经把每个Token的KV缓存大小压缩了54倍实在太惊人了所以在今天的视频里我们就来看看DeepSeek是如何降低前沿AI模型成本的以及他们想" } ], "id": 20, "review_annotation": "", "row_id": 20, "bounded_row_id": 20, "parent_row_ids": [ 20 ], "parent_row_id": 20, "packet_id": 1, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 26.787985576923077, "packet_end": 55.97597115384615, "raw_text": "实在太惊人了", "selected_text": "实在太惊人了", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 20 ], "packet_id": 1, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 49.70798557692308, "end": 51.24798557692308, "text": "所以在今天的视频里", "raw_asr_text": "所以在今天的视频里", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 1, "words": [ { "word": "所以", "start": 49.70798557692308, "end": 50.38798557692308, "probability": 0.93310546875 }, { "word": "在", "start": 50.38798557692308, "end": 50.52798557692307, "probability": 0.9921875 }, { "word": "今天的", "start": 50.52798557692307, "end": 50.78798557692308, "probability": 0.97021484375 }, { "word": "视", "start": 50.78798557692308, "end": 50.98798557692308, "probability": 0.9990234375 }, { "word": "频", "start": 50.98798557692308, "end": 51.10798557692308, "probability": 0.99951171875 }, { "word": "里", "start": 51.10798557692308, "end": 51.24798557692308, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.16982789856294242, "compression_ratio": 1.1712538226299694, "no_speech_prob": 5.825580295937272e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "更疯狂V4.1Flash的参数量几乎是V4Flash的2倍从2840亿增长到5520亿而且其新架构还做了多方面精简足以把KV缓存再压缩4倍为了让你明白这有多厉害过去9个月里DeepSeek已经把每个Token的KV缓存大小压缩了54倍实在太惊人了所以在今天的视频里我们就来看看DeepSeek是如何降低前沿AI模型成本的以及他们想" } ], "id": 21, "review_annotation": "", "row_id": 21, "bounded_row_id": 21, "parent_row_ids": [ 21 ], "parent_row_id": 21, "packet_id": 1, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 26.787985576923077, "packet_end": 55.97597115384615, "raw_text": "所以在今天的视频里", "selected_text": "所以在今天的视频里", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 21 ], "packet_id": 1, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 51.24798557692308, "end": 52.10798557692308, "text": "我们就来看看", "raw_asr_text": "我们就来看看", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 1, "words": [ { "word": "我们", "start": 51.24798557692308, "end": 51.56798557692308, "probability": 0.99462890625 }, { "word": "就", "start": 51.56798557692308, "end": 51.687985576923076, "probability": 0.99267578125 }, { "word": "来", "start": 51.687985576923076, "end": 51.80798557692307, "probability": 0.99853515625 }, { "word": "看看", "start": 51.80798557692307, "end": 52.10798557692308, "probability": 0.9755859375 } ], "asr_metrics": { "avg_logprob": -0.16982789856294242, "compression_ratio": 1.1712538226299694, "no_speech_prob": 5.825580295937272e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "更疯狂V4.1Flash的参数量几乎是V4Flash的2倍从2840亿增长到5520亿而且其新架构还做了多方面精简足以把KV缓存再压缩4倍为了让你明白这有多厉害过去9个月里DeepSeek已经把每个Token的KV缓存大小压缩了54倍实在太惊人了所以在今天的视频里我们就来看看DeepSeek是如何降低前沿AI模型成本的以及他们想" } ], "id": 22, "review_annotation": "", "row_id": 22, "bounded_row_id": 22, "parent_row_ids": [ 22 ], "parent_row_id": 22, "packet_id": 1, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 26.787985576923077, "packet_end": 55.97597115384615, "raw_text": "我们就来看看", "selected_text": "我们就来看看", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 22 ], "packet_id": 1, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 52.10798557692308, "end": 55.02798557692307, "text": "DeepSeek是如何降低前沿AI模型成本的", "raw_asr_text": "DeepSeek是如何降低前沿AI模型成本的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 1, "words": [ { "word": "De", "start": 52.10798557692308, "end": 52.34798557692308, "probability": 0.98583984375 }, { "word": "ep", "start": 52.34798557692308, "end": 52.42798557692308, "probability": 0.9990234375 }, { "word": "Se", "start": 52.42798557692308, "end": 52.56798557692308, "probability": 0.9755859375 }, { "word": "ek", "start": 52.56798557692308, "end": 52.66798557692307, "probability": 1.0 }, { "word": "是", "start": 52.66798557692307, "end": 52.84798557692308, "probability": 0.9873046875 }, { "word": "如何", "start": 52.84798557692308, "end": 53.14798557692308, "probability": 0.99658203125 }, { "word": "降", "start": 53.14798557692308, "end": 53.367985576923076, "probability": 0.9873046875 }, { "word": "低", "start": 53.367985576923076, "end": 53.56798557692308, "probability": 0.9951171875 }, { "word": "前", "start": 53.56798557692308, "end": 53.78798557692308, "probability": 0.9501953125 }, { "word": "沿", "start": 53.78798557692308, "end": 53.96798557692308, "probability": 0.998046875 }, { "word": "AI", "start": 53.96798557692308, "end": 54.24798557692308, "probability": 0.96728515625 }, { "word": "模", "start": 54.24798557692308, "end": 54.40798557692308, "probability": 0.9892578125 }, { "word": "型", "start": 54.40798557692308, "end": 54.54798557692308, "probability": 0.9990234375 }, { "word": "成", "start": 54.54798557692308, "end": 54.687985576923076, "probability": 0.9912109375 }, { "word": "本", "start": 54.687985576923076, "end": 54.867985576923076, "probability": 0.99951171875 }, { "word": "的", "start": 54.867985576923076, "end": 55.02798557692307, "probability": 0.99658203125 } ], "asr_metrics": { "avg_logprob": -0.16982789856294242, "compression_ratio": 1.1712538226299694, "no_speech_prob": 5.825580295937272e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "更疯狂V4.1Flash的参数量几乎是V4Flash的2倍从2840亿增长到5520亿而且其新架构还做了多方面精简足以把KV缓存再压缩4倍为了让你明白这有多厉害过去9个月里DeepSeek已经把每个Token的KV缓存大小压缩了54倍实在太惊人了所以在今天的视频里我们就来看看DeepSeek是如何降低前沿AI模型成本的以及他们想" } ], "id": 23, "review_annotation": "", "row_id": 23, "bounded_row_id": 23, "parent_row_ids": [ 23 ], "parent_row_id": 23, "packet_id": 1, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 26.787985576923077, "packet_end": 55.97597115384615, "raw_text": "DeepSeek是如何降低前沿AI模型成本的", "selected_text": "DeepSeek是如何降低前沿AI模型成本的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 23 ], "packet_id": 1, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 55.175971153846156, "end": 57.59597115384616, "text": "以及他们想出了哪些疯狂的新技术", "raw_asr_text": "以及他们想出了哪些疯狂的新技术", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "以及", "start": 55.13597115384616, "end": 55.45597115384616, "probability": 0.1627197265625 }, { "word": "他们", "start": 55.45597115384616, "end": 55.75597115384616, "probability": 0.76806640625 }, { "word": "想", "start": 55.75597115384616, "end": 55.93597115384616, "probability": 0.99755859375 }, { "word": "出", "start": 55.93597115384616, "end": 56.09597115384616, "probability": 0.99560546875 }, { "word": "了", "start": 56.09597115384616, "end": 56.175971153846156, "probability": 0.97998046875 }, { "word": "哪", "start": 56.175971153846156, "end": 56.33597115384616, "probability": 0.955078125 }, { "word": "些", "start": 56.33597115384616, "end": 56.53597115384616, "probability": 0.9990234375 }, { "word": "疯", "start": 56.53597115384616, "end": 56.77597115384616, "probability": 0.9970703125 }, { "word": "狂", "start": 56.77597115384616, "end": 56.95597115384616, "probability": 0.9990234375 }, { "word": "的", "start": 56.95597115384616, "end": 57.05597115384616, "probability": 0.99560546875 }, { "word": "新", "start": 57.05597115384616, "end": 57.215971153846155, "probability": 0.99267578125 }, { "word": "技", "start": 57.215971153846155, "end": 57.39597115384616, "probability": 0.99853515625 }, { "word": "术", "start": 57.39597115384616, "end": 57.59597115384616, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 24, "review_annotation": "", "row_id": 24, "bounded_row_id": 24, "parent_row_ids": [ 24 ], "parent_row_id": 24, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "以及他们想出了哪些疯狂的新技术", "selected_text": "以及他们想出了哪些疯狂的新技术", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 24 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 57.59597115384616, "end": 59.15597115384616, "text": "才让这一切成为可能", "raw_asr_text": "才让这一切成为可能", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "才", "start": 57.59597115384616, "end": 57.89597115384616, "probability": 0.86767578125 }, { "word": "让", "start": 57.89597115384616, "end": 58.03597115384616, "probability": 0.98876953125 }, { "word": "这", "start": 58.03597115384616, "end": 58.13597115384616, "probability": 0.95947265625 }, { "word": "一", "start": 58.13597115384616, "end": 58.31597115384616, "probability": 0.99365234375 }, { "word": "切", "start": 58.31597115384616, "end": 58.47597115384616, "probability": 0.99951171875 }, { "word": "成", "start": 58.47597115384616, "end": 58.675971153846156, "probability": 0.99755859375 }, { "word": "为", "start": 58.675971153846156, "end": 58.81597115384616, "probability": 0.99609375 }, { "word": "可能", "start": 58.81597115384616, "end": 59.15597115384616, "probability": 0.99560546875 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 25, "review_annotation": "", "row_id": 25, "bounded_row_id": 25, "parent_row_ids": [ 25 ], "parent_row_id": 25, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "才让这一切成为可能", "selected_text": "才让这一切成为可能", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 25 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 59.15597115384616, "end": 60.47597115384616, "text": "总之回到视频正题", "raw_asr_text": "总之回到视频正题", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "总", "start": 59.15597115384616, "end": 59.39597115384616, "probability": 0.94677734375 }, { "word": "之", "start": 59.39597115384616, "end": 59.57597115384616, "probability": 0.99609375 }, { "word": "回", "start": 59.57597115384616, "end": 59.81597115384616, "probability": 0.5771484375 }, { "word": "到", "start": 59.81597115384616, "end": 59.91597115384616, "probability": 0.9990234375 }, { "word": "视", "start": 59.91597115384616, "end": 60.05597115384616, "probability": 0.9970703125 }, { "word": "频", "start": 60.05597115384616, "end": 60.13597115384616, "probability": 0.99951171875 }, { "word": "正", "start": 60.13597115384616, "end": 60.27597115384616, "probability": 0.98486328125 }, { "word": "题", "start": 60.27597115384616, "end": 60.47597115384616, "probability": 0.99658203125 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 26, "review_annotation": "", "row_id": 26, "bounded_row_id": 26, "parent_row_ids": [ 26 ], "parent_row_id": 26, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "总之回到视频正题", "selected_text": "总之回到视频正题", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 26 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 60.47597115384616, "end": 61.995971153846156, "text": "现在我觉得很明显", "raw_asr_text": "现在我觉得很明显", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "现在", "start": 60.47597115384616, "end": 60.91597115384616, "probability": 0.94580078125 }, { "word": "我觉得", "start": 60.91597115384616, "end": 61.39597115384616, "probability": 0.9658203125 }, { "word": "很", "start": 61.39597115384616, "end": 61.63597115384616, "probability": 0.99267578125 }, { "word": "明", "start": 61.63597115384616, "end": 61.79597115384616, "probability": 0.99951171875 }, { "word": "显", "start": 61.79597115384616, "end": 61.995971153846156, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 27, "review_annotation": "", "row_id": 27, "bounded_row_id": 27, "parent_row_ids": [ 27 ], "parent_row_id": 27, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "现在我觉得很明显", "selected_text": "现在我觉得很明显", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 27 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 61.995971153846156, "end": 63.39597115384616, "text": "Deep Seek并没有追逐", "raw_asr_text": "Deep Seek并没有追逐", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "De", "start": 61.995971153846156, "end": 62.19597115384616, "probability": 0.74609375 }, { "word": "ep", "start": 62.19597115384616, "end": 62.27597115384616, "probability": 0.99072265625 }, { "word": " Se", "start": 62.27597115384616, "end": 62.41597115384616, "probability": 0.3291015625 }, { "word": "ek", "start": 62.41597115384616, "end": 62.51597115384616, "probability": 0.97412109375 }, { "word": "并", "start": 62.51597115384616, "end": 62.73597115384616, "probability": 0.97265625 }, { "word": "没有", "start": 62.73597115384616, "end": 62.95597115384616, "probability": 0.99658203125 }, { "word": "追", "start": 62.95597115384616, "end": 63.19597115384616, "probability": 0.99755859375 }, { "word": "逐", "start": 63.19597115384616, "end": 63.39597115384616, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 28, "review_annotation": "", "row_id": 28, "bounded_row_id": 28, "parent_row_ids": [ 28 ], "parent_row_id": 28, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "Deep Seek并没有追逐", "selected_text": "Deep Seek并没有追逐", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 28 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 63.39597115384616, "end": 65.15597115384615, "text": "基准测试中的最先进地位", "raw_asr_text": "基准测试中的最先进地位", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "基", "start": 63.39597115384616, "end": 63.59597115384616, "probability": 0.89990234375 }, { "word": "准", "start": 63.59597115384616, "end": 63.75597115384616, "probability": 0.9912109375 }, { "word": "测", "start": 63.75597115384616, "end": 63.89597115384616, "probability": 0.9990234375 }, { "word": "试", "start": 63.89597115384616, "end": 64.05597115384616, "probability": 0.99853515625 }, { "word": "中", "start": 64.05597115384616, "end": 64.15597115384615, "probability": 0.98876953125 }, { "word": "的", "start": 64.15597115384615, "end": 64.31597115384616, "probability": 0.9931640625 }, { "word": "最", "start": 64.31597115384616, "end": 64.49597115384616, "probability": 0.998046875 }, { "word": "先", "start": 64.49597115384616, "end": 64.67597115384616, "probability": 0.9990234375 }, { "word": "进", "start": 64.67597115384616, "end": 64.79597115384615, "probability": 0.9990234375 }, { "word": "地", "start": 64.79597115384615, "end": 64.95597115384616, "probability": 0.9921875 }, { "word": "位", "start": 64.95597115384616, "end": 65.15597115384615, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 29, "review_annotation": "", "row_id": 29, "bounded_row_id": 29, "parent_row_ids": [ 29 ], "parent_row_id": 29, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "基准测试中的最先进地位", "selected_text": "基准测试中的最先进地位", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 29 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 65.15597115384615, "end": 67.47597115384616, "text": "而是把实际运行大元模型这件事", "raw_asr_text": "而是把实际运行大元模型这件事", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "而", "start": 65.15597115384615, "end": 65.41597115384616, "probability": 0.990234375 }, { "word": "是", "start": 65.41597115384616, "end": 65.53597115384616, "probability": 0.9990234375 }, { "word": "把", "start": 65.53597115384616, "end": 65.67597115384616, "probability": 0.9970703125 }, { "word": "实", "start": 65.67597115384616, "end": 65.95597115384616, "probability": 0.998046875 }, { "word": "际", "start": 65.95597115384616, "end": 66.13597115384616, "probability": 1.0 }, { "word": "运", "start": 66.13597115384616, "end": 66.33597115384616, "probability": 0.9990234375 }, { "word": "行", "start": 66.33597115384616, "end": 66.47597115384616, "probability": 0.99267578125 }, { "word": "大", "start": 66.47597115384616, "end": 66.67597115384616, "probability": 0.978515625 }, { "word": "元", "start": 66.67597115384616, "end": 66.83597115384616, "probability": 0.451904296875 }, { "word": "模", "start": 66.83597115384616, "end": 66.97597115384616, "probability": 0.994140625 }, { "word": "型", "start": 66.97597115384616, "end": 67.09597115384616, "probability": 0.9970703125 }, { "word": "这", "start": 67.09597115384616, "end": 67.21597115384616, "probability": 0.98486328125 }, { "word": "件", "start": 67.21597115384616, "end": 67.31597115384616, "probability": 0.9970703125 }, { "word": "事", "start": 67.31597115384616, "end": 67.47597115384616, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 30, "review_annotation": "", "row_id": 30, "bounded_row_id": 30, "parent_row_ids": [ 30 ], "parent_row_id": 30, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "而是把实际运行大元模型这件事", "selected_text": "而是把实际运行大元模型这件事", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 30 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 67.47597115384616, "end": 68.47597115384616, "text": "优化到了极致", "raw_asr_text": "优化到了极致", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "优", "start": 67.47597115384616, "end": 67.71597115384616, "probability": 0.9560546875 }, { "word": "化", "start": 67.71597115384616, "end": 67.87597115384617, "probability": 0.99951171875 }, { "word": "到了", "start": 67.87597115384617, "end": 68.07597115384615, "probability": 0.97900390625 }, { "word": "极", "start": 68.07597115384615, "end": 68.31597115384616, "probability": 0.9951171875 }, { "word": "致", "start": 68.31597115384616, "end": 68.47597115384616, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 31, "review_annotation": "", "row_id": 31, "bounded_row_id": 31, "parent_row_ids": [ 31 ], "parent_row_id": 31, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "优化到了极致", "selected_text": "优化到了极致", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 31 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 68.47597115384616, "end": 70.93597115384615, "text": "你多少能从性能跃升中看出这一点", "raw_asr_text": "你多少能从性能跃升中看出这一点", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "你", "start": 68.47597115384616, "end": 68.83597115384616, "probability": 0.9775390625 }, { "word": "多少", "start": 68.83597115384616, "end": 69.11597115384616, "probability": 0.99609375 }, { "word": "能", "start": 69.11597115384616, "end": 69.29597115384615, "probability": 0.9873046875 }, { "word": "从", "start": 69.29597115384615, "end": 69.43597115384615, "probability": 0.9990234375 }, { "word": "性", "start": 69.43597115384615, "end": 69.63597115384616, "probability": 0.9970703125 }, { "word": "能", "start": 69.63597115384616, "end": 69.79597115384615, "probability": 0.99951171875 }, { "word": "跃", "start": 69.79597115384615, "end": 69.91597115384616, "probability": 0.8603515625 }, { "word": "升", "start": 69.91597115384616, "end": 70.05597115384616, "probability": 0.98779296875 }, { "word": "中", "start": 70.05597115384616, "end": 70.17597115384616, "probability": 0.99462890625 }, { "word": "看", "start": 70.17597115384616, "end": 70.45597115384615, "probability": 0.99365234375 }, { "word": "出", "start": 70.45597115384615, "end": 70.59597115384616, "probability": 0.9931640625 }, { "word": "这", "start": 70.59597115384616, "end": 70.73597115384615, "probability": 0.99658203125 }, { "word": "一点", "start": 70.73597115384615, "end": 70.93597115384615, "probability": 0.9853515625 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 32, "review_annotation": "", "row_id": 32, "bounded_row_id": 32, "parent_row_ids": [ 32 ], "parent_row_id": 32, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "你多少能从性能跃升中看出这一点", "selected_text": "你多少能从性能跃升中看出这一点", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 32 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 70.93597115384615, "end": 71.85597115384616, "text": "而事后看来", "raw_asr_text": "而事后看来", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "而", "start": 70.93597115384615, "end": 71.21597115384616, "probability": 0.96875 }, { "word": "事", "start": 71.21597115384616, "end": 71.41597115384616, "probability": 0.990234375 }, { "word": "后", "start": 71.41597115384616, "end": 71.55597115384616, "probability": 0.99755859375 }, { "word": "看", "start": 71.55597115384616, "end": 71.69597115384616, "probability": 0.99267578125 }, { "word": "来", "start": 71.69597115384616, "end": 71.85597115384616, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 33, "review_annotation": "", "row_id": 33, "bounded_row_id": 33, "parent_row_ids": [ 33 ], "parent_row_id": 33, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "而事后看来", "selected_text": "而事后看来", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 33 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 71.85597115384616, "end": 73.61597115384616, "text": "这种跃升也并不算特别疯狂", "raw_asr_text": "这种跃升也并不算特别疯狂", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "这", "start": 71.85597115384616, "end": 72.07597115384615, "probability": 0.9755859375 }, { "word": "种", "start": 72.07597115384615, "end": 72.13597115384616, "probability": 0.9990234375 }, { "word": "跃", "start": 72.13597115384616, "end": 72.27597115384616, "probability": 0.9990234375 }, { "word": "升", "start": 72.27597115384616, "end": 72.43597115384615, "probability": 0.9990234375 }, { "word": "也", "start": 72.43597115384615, "end": 72.59597115384616, "probability": 0.9873046875 }, { "word": "并", "start": 72.59597115384616, "end": 72.79597115384615, "probability": 0.99365234375 }, { "word": "不", "start": 72.79597115384615, "end": 72.91597115384616, "probability": 0.99365234375 }, { "word": "算", "start": 72.91597115384616, "end": 73.01597115384615, "probability": 0.99658203125 }, { "word": "特", "start": 73.01597115384615, "end": 73.15597115384617, "probability": 0.990234375 }, { "word": "别", "start": 73.15597115384617, "end": 73.27597115384616, "probability": 0.998046875 }, { "word": "疯", "start": 73.27597115384616, "end": 73.43597115384615, "probability": 0.99951171875 }, { "word": "狂", "start": 73.43597115384615, "end": 73.61597115384616, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 34, "review_annotation": "", "row_id": 34, "bounded_row_id": 34, "parent_row_ids": [ 34 ], "parent_row_id": 34, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "这种跃升也并不算特别疯狂", "selected_text": "这种跃升也并不算特别疯狂", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 34 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 73.61597115384616, "end": 74.85597115384616, "text": "但当你进一步细看", "raw_asr_text": "但当你进一步细看", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "但", "start": 73.61597115384616, "end": 73.85597115384616, "probability": 0.97119140625 }, { "word": "当", "start": 73.85597115384616, "end": 74.01597115384615, "probability": 0.99658203125 }, { "word": "你", "start": 74.01597115384615, "end": 74.17597115384616, "probability": 0.998046875 }, { "word": "进", "start": 74.17597115384616, "end": 74.31597115384616, "probability": 0.9921875 }, { "word": "一", "start": 74.31597115384616, "end": 74.43597115384615, "probability": 0.99609375 }, { "word": "步", "start": 74.43597115384615, "end": 74.53597115384616, "probability": 0.99853515625 }, { "word": "细", "start": 74.53597115384616, "end": 74.69597115384616, "probability": 0.9345703125 }, { "word": "看", "start": 74.69597115384616, "end": 74.85597115384616, "probability": 0.935546875 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 35, "review_annotation": "", "row_id": 35, "bounded_row_id": 35, "parent_row_ids": [ 35 ], "parent_row_id": 35, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "但当你进一步细看", "selected_text": "但当你进一步细看", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 35 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 74.85597115384616, "end": 76.77597115384616, "text": "他们创造并落地的所有创新", "raw_asr_text": "他们创造并落地的所有创新", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "他们", "start": 74.85597115384616, "end": 75.03597115384616, "probability": 0.9462890625 }, { "word": "创", "start": 75.03597115384616, "end": 75.29597115384615, "probability": 0.9970703125 }, { "word": "造", "start": 75.29597115384615, "end": 75.47597115384616, "probability": 0.99853515625 }, { "word": "并", "start": 75.47597115384616, "end": 75.69597115384616, "probability": 0.99609375 }, { "word": "落", "start": 75.69597115384616, "end": 75.83597115384616, "probability": 0.99853515625 }, { "word": "地", "start": 75.83597115384616, "end": 75.97597115384616, "probability": 0.9892578125 }, { "word": "的", "start": 75.97597115384616, "end": 76.09597115384616, "probability": 0.99267578125 }, { "word": "所有", "start": 76.09597115384616, "end": 76.33597115384616, "probability": 0.99755859375 }, { "word": "创", "start": 76.33597115384616, "end": 76.57597115384615, "probability": 0.9990234375 }, { "word": "新", "start": 76.57597115384615, "end": 76.77597115384616, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 36, "review_annotation": "", "row_id": 36, "bounded_row_id": 36, "parent_row_ids": [ 36 ], "parent_row_id": 36, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "他们创造并落地的所有创新", "selected_text": "他们创造并落地的所有创新", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 36 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 76.77597115384616, "end": 78.97597115384616, "text": "尤其是考虑到模型规模后的", "raw_asr_text": "尤其是考虑到模型规模后的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "尤", "start": 76.77597115384616, "end": 77.31597115384616, "probability": 0.994140625 }, { "word": "其", "start": 77.31597115384616, "end": 77.41597115384616, "probability": 0.99951171875 }, { "word": "是", "start": 77.41597115384616, "end": 77.55597115384616, "probability": 0.99658203125 }, { "word": "考", "start": 77.55597115384616, "end": 77.71597115384616, "probability": 0.99755859375 }, { "word": "虑", "start": 77.71597115384616, "end": 77.85597115384616, "probability": 0.9990234375 }, { "word": "到", "start": 77.85597115384616, "end": 77.97597115384616, "probability": 0.99365234375 }, { "word": "模", "start": 77.97597115384616, "end": 78.17597115384616, "probability": 0.99462890625 }, { "word": "型", "start": 78.17597115384616, "end": 78.35597115384616, "probability": 1.0 }, { "word": "规", "start": 78.35597115384616, "end": 78.51597115384615, "probability": 0.99609375 }, { "word": "模", "start": 78.51597115384615, "end": 78.65597115384617, "probability": 0.99951171875 }, { "word": "后", "start": 78.65597115384617, "end": 78.81597115384616, "probability": 0.9755859375 }, { "word": "的", "start": 78.81597115384616, "end": 78.97597115384616, "probability": 0.9921875 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 37, "review_annotation": "", "row_id": 37, "bounded_row_id": 37, "parent_row_ids": [ 37 ], "parent_row_id": 37, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "尤其是考虑到模型规模后的", "selected_text": "尤其是考虑到模型规模后的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 37 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 78.97597115384616, "end": 80.53597115384616, "text": "每百万Token价格", "raw_asr_text": "每百万Token价格", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "每", "start": 78.97597115384616, "end": 79.23597115384615, "probability": 0.99267578125 }, { "word": "百", "start": 79.23597115384615, "end": 79.53597115384616, "probability": 0.9736328125 }, { "word": "万", "start": 79.53597115384616, "end": 79.75597115384616, "probability": 0.9990234375 }, { "word": "T", "start": 79.75597115384616, "end": 79.93597115384615, "probability": 0.62890625 }, { "word": "oken", "start": 79.93597115384615, "end": 80.09597115384616, "probability": 0.576171875 }, { "word": "价", "start": 80.09597115384616, "end": 80.35597115384616, "probability": 0.7900390625 }, { "word": "格", "start": 80.35597115384616, "end": 80.53597115384616, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 38, "review_annotation": "", "row_id": 38, "bounded_row_id": 38, "parent_row_ids": [ 38 ], "parent_row_id": 38, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "每百万Token价格", "selected_text": "每百万Token价格", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 38 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 80.53597115384616, "end": 82.03597115384616, "text": "他们无疑正在建立一种", "raw_asr_text": "他们无疑正在建立一种", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 2, "words": [ { "word": "他们", "start": 80.53597115384616, "end": 80.95597115384615, "probability": 0.98193359375 }, { "word": "无", "start": 80.95597115384615, "end": 81.13597115384616, "probability": 0.9990234375 }, { "word": "疑", "start": 81.13597115384616, "end": 81.29597115384615, "probability": 0.9931640625 }, { "word": "正", "start": 81.29597115384615, "end": 81.43597115384615, "probability": 0.99560546875 }, { "word": "在", "start": 81.43597115384615, "end": 81.57597115384615, "probability": 0.99853515625 }, { "word": "建", "start": 81.57597115384615, "end": 81.71597115384616, "probability": 0.99853515625 }, { "word": "立", "start": 81.71597115384616, "end": 81.81597115384616, "probability": 0.96240234375 }, { "word": "一", "start": 81.81597115384616, "end": 81.91597115384616, "probability": 0.98583984375 }, { "word": "种", "start": 81.91597115384616, "end": 82.03597115384616, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.11031976487901475, "compression_ratio": 1.2918781725888324, "no_speech_prob": 1.117682751905491e-10 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "以及他们想出了哪些疯狂的新技术才让这一切成为可能总之回到视频正题现在我觉得很明显Deep Seek并没有追逐基准测试中的最先进地位而是把实际运行大元模型这件事优化到了极致你多少能从性能跃升中看出这一点而事后看来这种跃升也并不算特别疯狂但当你进一步细看他们创造并落地的所有创新尤其是考虑到模型规模后的每百万Token价格他们无疑正在建立一种其他实验室望尘莫及的" } ], "id": 39, "review_annotation": "", "row_id": 39, "bounded_row_id": 39, "parent_row_ids": [ 39 ], "parent_row_id": 39, "packet_id": 2, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 54.37597115384616, "packet_end": 83.56395673076923, "raw_text": "他们无疑正在建立一种", "selected_text": "他们无疑正在建立一种", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 39 ], "packet_id": 2, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 82.76395673076924, "end": 84.38395673076924, "text": "其他实验室望尘莫及的惊人优势", "raw_asr_text": "其他实验室望尘莫及的惊人优势", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 3, "words": [ { "word": "其", "start": 81.96395673076924, "end": 82.26395673076924, "probability": 0.9404296875 }, { "word": "他", "start": 82.26395673076924, "end": 82.38395673076924, "probability": 0.8544921875 }, { "word": "实", "start": 82.38395673076924, "end": 82.56395673076923, "probability": 0.94580078125 }, { "word": "验", "start": 82.56395673076923, "end": 82.70395673076924, "probability": 0.9990234375 }, { "word": "室", "start": 82.70395673076924, "end": 82.92395673076923, "probability": 0.98779296875 }, { "word": "望", "start": 82.92395673076923, "end": 83.10395673076924, "probability": 0.63623046875 }, { "word": "尘", "start": 83.10395673076924, "end": 83.24395673076924, "probability": 0.955078125 }, { "word": "莫", "start": 83.24395673076924, "end": 83.40395673076924, "probability": 0.974609375 }, { "word": "及", "start": 83.40395673076924, "end": 83.52395673076924, "probability": 0.91943359375 }, { "word": "的", "start": 83.52395673076924, "end": 83.64395673076925, "probability": 0.9912109375 }, { "word": "惊", "start": 83.64395673076925, "end": 83.84395673076924, "probability": 0.99755859375 }, { "word": "人", "start": 83.84395673076924, "end": 83.98395673076924, "probability": 0.99951171875 }, { "word": "优", "start": 83.98395673076924, "end": 84.18395673076924, "probability": 0.998046875 }, { "word": "势", "start": 84.18395673076924, "end": 84.38395673076924, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.1741250594032002, "compression_ratio": 1.13946587537092, "no_speech_prob": 2.7717350983835765e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "其他实验室望尘莫及的惊人优势而当你站在企业的角度想一想如果一个模型质量能达到九成成本却只要1%你大概率会选这个方案对吧再往高了吹点在TerminalBench2.1上它的成本仅为Opus 5的5%是GPT5.6艘的5.8%是KIMI K3的8%同时性能还超过了这三款模型这就是Deep Seek的优势所在这是难度高得多的竞争赛道" } ], "id": 40, "review_annotation": "", "row_id": 40, "bounded_row_id": 40, "parent_row_ids": [ 40 ], "parent_row_id": 40, "packet_id": 3, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 81.96395673076924, "packet_end": 111.15194230769231, "raw_text": "其他实验室望尘莫及的惊人优势", "selected_text": "其他实验室望尘莫及的惊人优势", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 40 ], "packet_id": 3, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 84.38395673076924, "end": 86.44395673076924, "text": "而当你站在企业的角度想一想", "raw_asr_text": "而当你站在企业的角度想一想", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 3, "words": [ { "word": "而", "start": 84.38395673076924, "end": 84.72395673076925, "probability": 0.94873046875 }, { "word": "当", "start": 84.72395673076925, "end": 84.84395673076924, "probability": 0.984375 }, { "word": "你", "start": 84.84395673076924, "end": 85.00395673076925, "probability": 0.9931640625 }, { "word": "站", "start": 85.00395673076925, "end": 85.12395673076924, "probability": 0.9990234375 }, { "word": "在", "start": 85.12395673076924, "end": 85.26395673076924, "probability": 0.9970703125 }, { "word": "企", "start": 85.26395673076924, "end": 85.54395673076924, "probability": 0.998046875 }, { "word": "业", "start": 85.54395673076924, "end": 85.66395673076924, "probability": 0.9990234375 }, { "word": "的", "start": 85.66395673076924, "end": 85.74395673076924, "probability": 0.9912109375 }, { "word": "角", "start": 85.74395673076924, "end": 85.86395673076925, "probability": 0.99951171875 }, { "word": "度", "start": 85.86395673076925, "end": 86.00395673076925, "probability": 1.0 }, { "word": "想", "start": 86.00395673076925, "end": 86.16395673076924, "probability": 0.98974609375 }, { "word": "一", "start": 86.16395673076924, "end": 86.30395673076924, "probability": 0.98974609375 }, { "word": "想", "start": 86.30395673076924, "end": 86.44395673076924, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1741250594032002, "compression_ratio": 1.13946587537092, "no_speech_prob": 2.7717350983835765e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "其他实验室望尘莫及的惊人优势而当你站在企业的角度想一想如果一个模型质量能达到九成成本却只要1%你大概率会选这个方案对吧再往高了吹点在TerminalBench2.1上它的成本仅为Opus 5的5%是GPT5.6艘的5.8%是KIMI K3的8%同时性能还超过了这三款模型这就是Deep Seek的优势所在这是难度高得多的竞争赛道" } ], "id": 41, "review_annotation": "", "row_id": 41, "bounded_row_id": 41, "parent_row_ids": [ 41 ], "parent_row_id": 41, "packet_id": 3, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 81.96395673076924, "packet_end": 111.15194230769231, "raw_text": "而当你站在企业的角度想一想", "selected_text": "而当你站在企业的角度想一想", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 41 ], "packet_id": 3, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 86.44395673076924, "end": 88.82395673076924, "text": "如果一个模型质量能达到九成", "raw_asr_text": "如果一个模型质量能达到九成", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 3, "words": [ { "word": "如果", "start": 86.44395673076924, "end": 86.92395673076923, "probability": 0.97705078125 }, { "word": "一个", "start": 86.92395673076923, "end": 87.16395673076924, "probability": 0.994140625 }, { "word": "模", "start": 87.16395673076924, "end": 87.32395673076924, "probability": 0.9990234375 }, { "word": "型", "start": 87.32395673076924, "end": 87.50395673076925, "probability": 0.9990234375 }, { "word": "质", "start": 87.50395673076925, "end": 87.68395673076924, "probability": 0.9853515625 }, { "word": "量", "start": 87.68395673076924, "end": 87.80395673076924, "probability": 1.0 }, { "word": "能", "start": 87.80395673076924, "end": 87.94395673076924, "probability": 0.9453125 }, { "word": "达", "start": 87.94395673076924, "end": 88.08395673076924, "probability": 0.9990234375 }, { "word": "到", "start": 88.08395673076924, "end": 88.22395673076925, "probability": 0.99658203125 }, { "word": "九", "start": 88.22395673076925, "end": 88.48395673076924, "probability": 0.53076171875 }, { "word": "成", "start": 88.48395673076924, "end": 88.82395673076924, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.1741250594032002, "compression_ratio": 1.13946587537092, "no_speech_prob": 2.7717350983835765e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "其他实验室望尘莫及的惊人优势而当你站在企业的角度想一想如果一个模型质量能达到九成成本却只要1%你大概率会选这个方案对吧再往高了吹点在TerminalBench2.1上它的成本仅为Opus 5的5%是GPT5.6艘的5.8%是KIMI K3的8%同时性能还超过了这三款模型这就是Deep Seek的优势所在这是难度高得多的竞争赛道" } ], "id": 42, "review_annotation": "", "row_id": 42, "bounded_row_id": 42, "parent_row_ids": [ 42 ], "parent_row_id": 42, "packet_id": 3, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 81.96395673076924, "packet_end": 111.15194230769231, "raw_text": "如果一个模型质量能达到九成", "selected_text": "如果一个模型质量能达到九成", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 42 ], "packet_id": 3, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 88.82395673076924, "end": 90.08395673076924, "text": "成本却只要1%", "raw_asr_text": "成本却只要1%", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 3, "words": [ { "word": "成", "start": 88.82395673076924, "end": 89.22395673076925, "probability": 0.99169921875 }, { "word": "本", "start": 89.22395673076925, "end": 89.40395673076924, "probability": 0.99951171875 }, { "word": "却", "start": 89.40395673076924, "end": 89.58395673076924, "probability": 0.9951171875 }, { "word": "只", "start": 89.58395673076924, "end": 89.72395673076925, "probability": 0.9970703125 }, { "word": "要", "start": 89.72395673076925, "end": 89.86395673076925, "probability": 0.998046875 }, { "word": "1", "start": 89.86395673076925, "end": 90.04395673076924, "probability": 0.34521484375 }, { "word": "%", "start": 90.04395673076924, "end": 90.08395673076924, "probability": 0.0013275146484375 } ], "asr_metrics": { "avg_logprob": -0.1741250594032002, "compression_ratio": 1.13946587537092, "no_speech_prob": 2.7717350983835765e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "其他实验室望尘莫及的惊人优势而当你站在企业的角度想一想如果一个模型质量能达到九成成本却只要1%你大概率会选这个方案对吧再往高了吹点在TerminalBench2.1上它的成本仅为Opus 5的5%是GPT5.6艘的5.8%是KIMI K3的8%同时性能还超过了这三款模型这就是Deep Seek的优势所在这是难度高得多的竞争赛道" } ], "id": 43, "review_annotation": "", "row_id": 43, "bounded_row_id": 43, "parent_row_ids": [ 43 ], "parent_row_id": 43, "packet_id": 3, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 81.96395673076924, "packet_end": 111.15194230769231, "raw_text": "成本却只要1%", "selected_text": "成本却只要1%", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 43 ], "packet_id": 3, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 90.50395673076925, "end": 92.46395673076924, "text": "你大概率会选这个方案对吧", "raw_asr_text": "你大概率会选这个方案对吧", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 3, "words": [ { "word": "你", "start": 90.50395673076925, "end": 90.86395673076925, "probability": 0.336181640625 }, { "word": "大概", "start": 90.86395673076925, "end": 91.18395673076924, "probability": 0.99462890625 }, { "word": "率", "start": 91.18395673076924, "end": 91.38395673076924, "probability": 0.99853515625 }, { "word": "会", "start": 91.38395673076924, "end": 91.52395673076924, "probability": 0.9931640625 }, { "word": "选", "start": 91.52395673076924, "end": 91.68395673076924, "probability": 0.99853515625 }, { "word": "这个", "start": 91.68395673076924, "end": 91.86395673076925, "probability": 0.9853515625 }, { "word": "方", "start": 91.86395673076925, "end": 92.02395673076924, "probability": 0.99658203125 }, { "word": "案", "start": 92.02395673076924, "end": 92.18395673076924, "probability": 0.99462890625 }, { "word": "对", "start": 92.18395673076924, "end": 92.30395673076924, "probability": 0.85498046875 }, { "word": "吧", "start": 92.30395673076924, "end": 92.46395673076924, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.1741250594032002, "compression_ratio": 1.13946587537092, "no_speech_prob": 2.7717350983835765e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "其他实验室望尘莫及的惊人优势而当你站在企业的角度想一想如果一个模型质量能达到九成成本却只要1%你大概率会选这个方案对吧再往高了吹点在TerminalBench2.1上它的成本仅为Opus 5的5%是GPT5.6艘的5.8%是KIMI K3的8%同时性能还超过了这三款模型这就是Deep Seek的优势所在这是难度高得多的竞争赛道" } ], "id": 44, "review_annotation": "", "row_id": 44, "bounded_row_id": 44, "parent_row_ids": [ 44 ], "parent_row_id": 44, "packet_id": 3, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 81.96395673076924, "packet_end": 111.15194230769231, "raw_text": "你大概率会选这个方案对吧", "selected_text": "你大概率会选这个方案对吧", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 44 ], "packet_id": 3, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 92.46395673076924, "end": 93.90395673076924, "text": "再往高了吹点", "raw_asr_text": "再往高了吹点", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 3, "words": [ { "word": "再", "start": 92.46395673076924, "end": 92.82395673076924, "probability": 0.8046875 }, { "word": "往", "start": 92.82395673076924, "end": 93.06395673076923, "probability": 0.9892578125 }, { "word": "高", "start": 93.06395673076923, "end": 93.28395673076923, "probability": 0.97900390625 }, { "word": "了", "start": 93.28395673076923, "end": 93.42395673076925, "probability": 0.98876953125 }, { "word": "吹", "start": 93.42395673076925, "end": 93.68395673076924, "probability": 0.9755859375 }, { "word": "点", "start": 93.68395673076924, "end": 93.90395673076924, "probability": 0.9970703125 } ], "asr_metrics": { "avg_logprob": -0.1741250594032002, "compression_ratio": 1.13946587537092, "no_speech_prob": 2.7717350983835765e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "其他实验室望尘莫及的惊人优势而当你站在企业的角度想一想如果一个模型质量能达到九成成本却只要1%你大概率会选这个方案对吧再往高了吹点在TerminalBench2.1上它的成本仅为Opus 5的5%是GPT5.6艘的5.8%是KIMI K3的8%同时性能还超过了这三款模型这就是Deep Seek的优势所在这是难度高得多的竞争赛道" } ], "id": 45, "review_annotation": "", "row_id": 45, "bounded_row_id": 45, "parent_row_ids": [ 45 ], "parent_row_id": 45, "packet_id": 3, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 81.96395673076924, "packet_end": 111.15194230769231, "raw_text": "再往高了吹点", "selected_text": "再往高了吹点", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 45 ], "packet_id": 3, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 93.90395673076924, "end": 95.80395673076924, "text": "在TerminalBench2.1上", "raw_asr_text": "在TerminalBench2.1上", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 3, "words": [ { "word": "在", "start": 93.90395673076924, "end": 94.28395673076923, "probability": 0.9765625 }, { "word": "T", "start": 94.28395673076923, "end": 94.54395673076924, "probability": 0.8388671875 }, { "word": "erm", "start": 94.54395673076924, "end": 94.64395673076925, "probability": 0.7841796875 }, { "word": "inal", "start": 94.64395673076925, "end": 94.78395673076923, "probability": 0.94580078125 }, { "word": "Ben", "start": 94.78395673076923, "end": 94.94395673076924, "probability": 0.496826171875 }, { "word": "ch", "start": 94.94395673076924, "end": 95.08395673076924, "probability": 0.998046875 }, { "word": "2.", "start": 95.08395673076924, "end": 95.24395673076924, "probability": 0.51171875 }, { "word": "1", "start": 95.36395673076925, "end": 95.58395673076924, "probability": 0.9970703125 }, { "word": "上", "start": 95.58395673076924, "end": 95.80395673076924, "probability": 0.9560546875 } ], "asr_metrics": { "avg_logprob": -0.1741250594032002, "compression_ratio": 1.13946587537092, "no_speech_prob": 2.7717350983835765e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "其他实验室望尘莫及的惊人优势而当你站在企业的角度想一想如果一个模型质量能达到九成成本却只要1%你大概率会选这个方案对吧再往高了吹点在TerminalBench2.1上它的成本仅为Opus 5的5%是GPT5.6艘的5.8%是KIMI K3的8%同时性能还超过了这三款模型这就是Deep Seek的优势所在这是难度高得多的竞争赛道" } ], "id": 46, "review_annotation": "", "row_id": 46, "bounded_row_id": 46, "parent_row_ids": [ 46 ], "parent_row_id": 46, "packet_id": 3, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 81.96395673076924, "packet_end": 111.15194230769231, "raw_text": "在TerminalBench2.1上", "selected_text": "在TerminalBench2.1上", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 46 ], "packet_id": 3, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 95.80395673076924, "end": 99.28395673076923, "text": "它的成本仅为Opus 5的5%", "raw_asr_text": "它的成本仅为Opus 5的5%", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 3, "words": [ { "word": "它的", "start": 95.80395673076924, "end": 96.16395673076924, "probability": 0.9306640625 }, { "word": "成", "start": 96.16395673076924, "end": 96.42395673076925, "probability": 0.99951171875 }, { "word": "本", "start": 96.42395673076925, "end": 96.70395673076924, "probability": 0.9990234375 }, { "word": "仅", "start": 96.70395673076924, "end": 97.12395673076924, "probability": 0.99755859375 }, { "word": "为", "start": 97.12395673076924, "end": 97.30395673076924, "probability": 0.99658203125 }, { "word": "O", "start": 97.30395673076924, "end": 97.44395673076924, "probability": 0.4775390625 }, { "word": "pus", "start": 97.44395673076924, "end": 97.64395673076925, "probability": 0.53466796875 }, { "word": " 5", "start": 97.64395673076925, "end": 97.92395673076925, "probability": 0.744140625 }, { "word": "的", "start": 97.92395673076925, "end": 98.10395673076924, "probability": 0.7705078125 }, { "word": "5", "start": 98.10395673076924, "end": 98.96395673076924, "probability": 0.98779296875 }, { "word": "%", "start": 98.96395673076924, "end": 99.28395673076923, "probability": 0.96142578125 } ], "asr_metrics": { "avg_logprob": -0.1741250594032002, "compression_ratio": 1.13946587537092, "no_speech_prob": 2.7717350983835765e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "其他实验室望尘莫及的惊人优势而当你站在企业的角度想一想如果一个模型质量能达到九成成本却只要1%你大概率会选这个方案对吧再往高了吹点在TerminalBench2.1上它的成本仅为Opus 5的5%是GPT5.6艘的5.8%是KIMI K3的8%同时性能还超过了这三款模型这就是Deep Seek的优势所在这是难度高得多的竞争赛道" } ], "id": 47, "review_annotation": "", "row_id": 47, "bounded_row_id": 47, "parent_row_ids": [ 47 ], "parent_row_id": 47, "packet_id": 3, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 81.96395673076924, "packet_end": 111.15194230769231, "raw_text": "它的成本仅为Opus 5的5%", "selected_text": "它的成本仅为Opus 5的5%", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 47 ], "packet_id": 3, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 99.28395673076923, "end": 102.24395673076924, "text": "是GPT5.6艘的5.8%", "raw_asr_text": "是GPT5.6艘的5.8%", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 3, "words": [ { "word": "是", "start": 99.28395673076923, "end": 99.54395673076924, "probability": 0.95556640625 }, { "word": "G", "start": 99.54395673076924, "end": 99.78395673076923, "probability": 0.95654296875 }, { "word": "P", "start": 99.78395673076923, "end": 99.88395673076924, "probability": 0.904296875 }, { "word": "T", "start": 99.88395673076924, "end": 100.02395673076924, "probability": 0.98486328125 }, { "word": "5.", "start": 100.02395673076924, "end": 100.24395673076924, "probability": 0.70751953125 }, { "word": "6", "start": 100.36395673076925, "end": 100.58395673076924, "probability": 0.98974609375 }, { "word": "艘", "start": 100.58395673076924, "end": 100.74395673076924, "probability": 0.55517578125 }, { "word": "的", "start": 100.74395673076924, "end": 100.92395673076925, "probability": 0.92919921875 }, { "word": "5.", "start": 100.92395673076925, "end": 101.16395673076924, "probability": 0.99609375 }, { "word": "8", "start": 101.72395673076925, "end": 101.96395673076924, "probability": 0.998046875 }, { "word": "%", "start": 101.96395673076924, "end": 102.24395673076924, "probability": 0.99365234375 } ], "asr_metrics": { "avg_logprob": -0.1741250594032002, "compression_ratio": 1.13946587537092, "no_speech_prob": 2.7717350983835765e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "其他实验室望尘莫及的惊人优势而当你站在企业的角度想一想如果一个模型质量能达到九成成本却只要1%你大概率会选这个方案对吧再往高了吹点在TerminalBench2.1上它的成本仅为Opus 5的5%是GPT5.6艘的5.8%是KIMI K3的8%同时性能还超过了这三款模型这就是Deep Seek的优势所在这是难度高得多的竞争赛道" } ], "id": 48, "review_annotation": "", "row_id": 48, "bounded_row_id": 48, "parent_row_ids": [ 48 ], "parent_row_id": 48, "packet_id": 3, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 81.96395673076924, "packet_end": 111.15194230769231, "raw_text": "是GPT5.6艘的5.8%", "selected_text": "是GPT5.6艘的5.8%", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 48 ], "packet_id": 3, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 102.24395673076924, "end": 104.44395673076924, "text": "是KIMI K3的8%", "raw_asr_text": "是KIMI K3的8%", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 3, "words": [ { "word": "是", "start": 102.24395673076924, "end": 102.42395673076925, "probability": 0.986328125 }, { "word": "K", "start": 102.42395673076925, "end": 102.68395673076924, "probability": 0.8974609375 }, { "word": "IM", "start": 102.68395673076924, "end": 102.86395673076925, "probability": 0.44189453125 }, { "word": "I", "start": 102.86395673076925, "end": 102.94395673076924, "probability": 0.80615234375 }, { "word": " K", "start": 102.94395673076924, "end": 103.10395673076924, "probability": 0.673828125 }, { "word": "3", "start": 103.10395673076924, "end": 103.34395673076924, "probability": 0.9873046875 }, { "word": "的", "start": 103.34395673076924, "end": 103.50395673076923, "probability": 0.98095703125 }, { "word": "8", "start": 103.50395673076923, "end": 103.68395673076924, "probability": 0.9951171875 }, { "word": "%", "start": 103.68395673076924, "end": 104.44395673076924, "probability": 0.99658203125 } ], "asr_metrics": { "avg_logprob": -0.1741250594032002, "compression_ratio": 1.13946587537092, "no_speech_prob": 2.7717350983835765e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "其他实验室望尘莫及的惊人优势而当你站在企业的角度想一想如果一个模型质量能达到九成成本却只要1%你大概率会选这个方案对吧再往高了吹点在TerminalBench2.1上它的成本仅为Opus 5的5%是GPT5.6艘的5.8%是KIMI K3的8%同时性能还超过了这三款模型这就是Deep Seek的优势所在这是难度高得多的竞争赛道" } ], "id": 49, "review_annotation": "", "row_id": 49, "bounded_row_id": 49, "parent_row_ids": [ 49 ], "parent_row_id": 49, "packet_id": 3, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 81.96395673076924, "packet_end": 111.15194230769231, "raw_text": "是KIMI K3的8%", "selected_text": "是KIMI K3的8%", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 49 ], "packet_id": 3, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 104.44395673076924, "end": 106.66395673076924, "text": "同时性能还超过了这三款模型", "raw_asr_text": "同时性能还超过了这三款模型", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 3, "words": [ { "word": "同", "start": 104.44395673076924, "end": 104.64395673076925, "probability": 0.98291015625 }, { "word": "时", "start": 104.64395673076925, "end": 104.92395673076925, "probability": 0.99951171875 }, { "word": "性", "start": 104.92395673076925, "end": 105.22395673076925, "probability": 0.90625 }, { "word": "能", "start": 105.22395673076925, "end": 105.36395673076925, "probability": 0.99853515625 }, { "word": "还", "start": 105.36395673076925, "end": 105.48395673076924, "probability": 0.99609375 }, { "word": "超", "start": 105.48395673076924, "end": 105.64395673076925, "probability": 0.9990234375 }, { "word": "过", "start": 105.64395673076925, "end": 105.80395673076924, "probability": 0.9990234375 }, { "word": "了", "start": 105.80395673076924, "end": 105.88395673076924, "probability": 0.95849609375 }, { "word": "这", "start": 105.88395673076924, "end": 106.04395673076924, "probability": 0.98046875 }, { "word": "三", "start": 106.04395673076924, "end": 106.20395673076924, "probability": 0.86865234375 }, { "word": "款", "start": 106.20395673076924, "end": 106.36395673076925, "probability": 0.99853515625 }, { "word": "模", "start": 106.36395673076925, "end": 106.48395673076924, "probability": 0.998046875 }, { "word": "型", "start": 106.48395673076924, "end": 106.66395673076924, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.1741250594032002, "compression_ratio": 1.13946587537092, "no_speech_prob": 2.7717350983835765e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "其他实验室望尘莫及的惊人优势而当你站在企业的角度想一想如果一个模型质量能达到九成成本却只要1%你大概率会选这个方案对吧再往高了吹点在TerminalBench2.1上它的成本仅为Opus 5的5%是GPT5.6艘的5.8%是KIMI K3的8%同时性能还超过了这三款模型这就是Deep Seek的优势所在这是难度高得多的竞争赛道" } ], "id": 50, "review_annotation": "", "row_id": 50, "bounded_row_id": 50, "parent_row_ids": [ 50 ], "parent_row_id": 50, "packet_id": 3, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 81.96395673076924, "packet_end": 111.15194230769231, "raw_text": "同时性能还超过了这三款模型", "selected_text": "同时性能还超过了这三款模型", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 50 ], "packet_id": 3, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 106.66395673076924, "end": 109.00395673076923, "text": "这就是Deep Seek的优势所在", "raw_asr_text": "这就是Deep Seek的优势所在", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 3, "words": [ { "word": "这", "start": 106.66395673076924, "end": 106.98395673076924, "probability": 0.9931640625 }, { "word": "就是", "start": 106.98395673076924, "end": 107.28395673076923, "probability": 0.99755859375 }, { "word": "De", "start": 107.28395673076923, "end": 107.58395673076924, "probability": 0.93408203125 }, { "word": "ep", "start": 107.58395673076924, "end": 107.68395673076924, "probability": 0.9755859375 }, { "word": " Se", "start": 107.68395673076924, "end": 107.84395673076924, "probability": 0.365966796875 }, { "word": "ek", "start": 107.84395673076924, "end": 107.96395673076924, "probability": 0.98583984375 }, { "word": "的", "start": 107.96395673076924, "end": 108.12395673076924, "probability": 0.9716796875 }, { "word": "优", "start": 108.12395673076924, "end": 108.42395673076925, "probability": 0.99853515625 }, { "word": "势", "start": 108.42395673076925, "end": 108.62395673076924, "probability": 1.0 }, { "word": "所", "start": 108.62395673076924, "end": 108.82395673076924, "probability": 0.99609375 }, { "word": "在", "start": 108.82395673076924, "end": 109.00395673076923, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1741250594032002, "compression_ratio": 1.13946587537092, "no_speech_prob": 2.7717350983835765e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "其他实验室望尘莫及的惊人优势而当你站在企业的角度想一想如果一个模型质量能达到九成成本却只要1%你大概率会选这个方案对吧再往高了吹点在TerminalBench2.1上它的成本仅为Opus 5的5%是GPT5.6艘的5.8%是KIMI K3的8%同时性能还超过了这三款模型这就是Deep Seek的优势所在这是难度高得多的竞争赛道" } ], "id": 51, "review_annotation": "", "row_id": 51, "bounded_row_id": 51, "parent_row_ids": [ 51 ], "parent_row_id": 51, "packet_id": 3, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 81.96395673076924, "packet_end": 111.15194230769231, "raw_text": "这就是Deep Seek的优势所在", "selected_text": "这就是Deep Seek的优势所在", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 51 ], "packet_id": 3, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 109.00395673076923, "end": 110.35194230769231, "text": "这是难度高得多的竞争赛道", "raw_asr_text": "这是难度高得多的竞争赛道", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 3, "words": [ { "word": "这是", "start": 109.00395673076923, "end": 109.26395673076924, "probability": 0.9814453125 }, { "word": "难", "start": 109.26395673076924, "end": 109.48395673076924, "probability": 0.99560546875 }, { "word": "度", "start": 109.48395673076924, "end": 109.72395673076925, "probability": 0.9990234375 }, { "word": "高", "start": 109.72395673076925, "end": 109.86395673076925, "probability": 0.998046875 }, { "word": "得", "start": 109.86395673076925, "end": 110.02395673076924, "probability": 0.794921875 }, { "word": "多", "start": 110.02395673076924, "end": 110.22395673076925, "probability": 0.9990234375 }, { "word": "的", "start": 110.22395673076925, "end": 110.38395673076924, "probability": 0.99755859375 }, { "word": "竞", "start": 110.38395673076924, "end": 110.60395673076924, "probability": 0.998046875 }, { "word": "争", "start": 110.60395673076924, "end": 110.74395673076924, "probability": 0.9990234375 }, { "word": "赛", "start": 110.74395673076924, "end": 110.88395673076924, "probability": 0.9990234375 }, { "word": "道", "start": 110.88395673076924, "end": 111.02395673076924, "probability": 0.99609375 } ], "asr_metrics": { "avg_logprob": -0.1741250594032002, "compression_ratio": 1.13946587537092, "no_speech_prob": 2.7717350983835765e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "其他实验室望尘莫及的惊人优势而当你站在企业的角度想一想如果一个模型质量能达到九成成本却只要1%你大概率会选这个方案对吧再往高了吹点在TerminalBench2.1上它的成本仅为Opus 5的5%是GPT5.6艘的5.8%是KIMI K3的8%同时性能还超过了这三款模型这就是Deep Seek的优势所在这是难度高得多的竞争赛道" } ], "id": 52, "review_annotation": "", "row_id": 52, "bounded_row_id": 52, "parent_row_ids": [ 52 ], "parent_row_id": 52, "packet_id": 3, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 81.96395673076924, "packet_end": 111.15194230769231, "raw_text": "这是难度高得多的竞争赛道", "selected_text": "这是难度高得多的竞争赛道", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 52 ], "packet_id": 3, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 111.03194230769232, "end": 113.27194230769231, "text": "因为做优化可比刷宝难多了", "raw_asr_text": "因为做优化可比刷宝难多了", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 4, "words": [ { "word": "因为", "start": 111.03194230769232, "end": 111.35194230769231, "probability": 0.93115234375 }, { "word": "做", "start": 111.35194230769231, "end": 111.57194230769231, "probability": 0.9931640625 }, { "word": "优", "start": 111.57194230769231, "end": 111.69194230769232, "probability": 0.9990234375 }, { "word": "化", "start": 111.69194230769232, "end": 111.87194230769231, "probability": 0.99951171875 }, { "word": "可", "start": 111.87194230769231, "end": 112.13194230769231, "probability": 0.96044921875 }, { "word": "比", "start": 112.13194230769231, "end": 112.27194230769231, "probability": 0.9990234375 }, { "word": "刷", "start": 112.27194230769231, "end": 112.55194230769231, "probability": 0.9990234375 }, { "word": "宝", "start": 112.55194230769231, "end": 112.73194230769232, "probability": 0.9794921875 }, { "word": "难", "start": 112.73194230769232, "end": 112.91194230769231, "probability": 0.99755859375 }, { "word": "多", "start": 112.91194230769231, "end": 113.13194230769231, "probability": 0.99755859375 }, { "word": "了", "start": 113.13194230769231, "end": 113.27194230769231, "probability": 0.98828125 } ], "asr_metrics": { "avg_logprob": -0.1259358913744267, "compression_ratio": 1.3447293447293447, "no_speech_prob": 5.938804309435497e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "高得多的竞争赛道因为做优化可比刷宝难多了在这款全新的V4.1 Flash模型中它们的核心思路基本就是如何在做大模型的同时大幅降低它的内存占用这是一个优化问题目标是降低大元模型的使用成本更具体的说是降低重度智能体场景的使用成本因为如今大元模型的大部分工作负载都需要反复读取塞满了工具输出和代码库的海量历史记录Deepsy基本是全新投入优化长上下" } ], "id": 53, "review_annotation": "", "row_id": 53, "bounded_row_id": 53, "parent_row_ids": [ 53 ], "parent_row_id": 53, "packet_id": 4, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 109.55194230769231, "packet_end": 138.7399278846154, "raw_text": "因为做优化可比刷宝难多了", "selected_text": "因为做优化可比刷宝难多了", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 53 ], "packet_id": 4, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 113.27194230769231, "end": 116.05194230769231, "text": "在这款全新的V4.1 Flash模型中", "raw_asr_text": "在这款全新的V4.1 Flash模型中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 4, "words": [ { "word": "在", "start": 113.27194230769231, "end": 113.59194230769232, "probability": 0.9755859375 }, { "word": "这", "start": 113.59194230769232, "end": 113.75194230769232, "probability": 0.994140625 }, { "word": "款", "start": 113.75194230769232, "end": 113.93194230769231, "probability": 0.99853515625 }, { "word": "全", "start": 113.93194230769231, "end": 114.21194230769231, "probability": 0.99560546875 }, { "word": "新", "start": 114.21194230769231, "end": 114.41194230769231, "probability": 0.9990234375 }, { "word": "的", "start": 114.41194230769231, "end": 114.59194230769232, "probability": 0.99267578125 }, { "word": "V", "start": 114.59194230769232, "end": 114.83194230769232, "probability": 0.66796875 }, { "word": "4.", "start": 114.83194230769232, "end": 115.01194230769231, "probability": 0.98779296875 }, { "word": "1", "start": 115.13194230769231, "end": 115.27194230769231, "probability": 0.91015625 }, { "word": " Flash", "start": 115.27194230769231, "end": 115.51194230769231, "probability": 0.357421875 }, { "word": "模", "start": 115.51194230769231, "end": 115.73194230769232, "probability": 0.90478515625 }, { "word": "型", "start": 115.73194230769232, "end": 115.89194230769232, "probability": 0.998046875 }, { "word": "中", "start": 115.89194230769232, "end": 116.05194230769231, "probability": 0.994140625 } ], "asr_metrics": { "avg_logprob": -0.1259358913744267, "compression_ratio": 1.3447293447293447, "no_speech_prob": 5.938804309435497e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "高得多的竞争赛道因为做优化可比刷宝难多了在这款全新的V4.1 Flash模型中它们的核心思路基本就是如何在做大模型的同时大幅降低它的内存占用这是一个优化问题目标是降低大元模型的使用成本更具体的说是降低重度智能体场景的使用成本因为如今大元模型的大部分工作负载都需要反复读取塞满了工具输出和代码库的海量历史记录Deepsy基本是全新投入优化长上下" } ], "id": 54, "review_annotation": "", "row_id": 54, "bounded_row_id": 54, "parent_row_ids": [ 54 ], "parent_row_id": 54, "packet_id": 4, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 109.55194230769231, "packet_end": 138.7399278846154, "raw_text": "在这款全新的V4.1 Flash模型中", "selected_text": "在这款全新的V4.1 Flash模型中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 54 ], "packet_id": 4, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 116.05194230769231, "end": 117.25194230769232, "text": "它们的核心思路", "raw_asr_text": "它们的核心思路", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 4, "words": [ { "word": "它", "start": 116.05194230769231, "end": 116.37194230769231, "probability": 0.465576171875 }, { "word": "们", "start": 116.37194230769231, "end": 116.49194230769231, "probability": 0.9990234375 }, { "word": "的", "start": 116.49194230769231, "end": 116.57194230769231, "probability": 0.9921875 }, { "word": "核", "start": 116.57194230769231, "end": 116.73194230769232, "probability": 0.9990234375 }, { "word": "心", "start": 116.73194230769232, "end": 116.89194230769232, "probability": 0.99951171875 }, { "word": "思", "start": 116.89194230769232, "end": 117.11194230769232, "probability": 0.99951171875 }, { "word": "路", "start": 117.11194230769232, "end": 117.25194230769232, "probability": 0.99658203125 } ], "asr_metrics": { "avg_logprob": -0.1259358913744267, "compression_ratio": 1.3447293447293447, "no_speech_prob": 5.938804309435497e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "高得多的竞争赛道因为做优化可比刷宝难多了在这款全新的V4.1 Flash模型中它们的核心思路基本就是如何在做大模型的同时大幅降低它的内存占用这是一个优化问题目标是降低大元模型的使用成本更具体的说是降低重度智能体场景的使用成本因为如今大元模型的大部分工作负载都需要反复读取塞满了工具输出和代码库的海量历史记录Deepsy基本是全新投入优化长上下" } ], "id": 55, "review_annotation": "", "row_id": 55, "bounded_row_id": 55, "parent_row_ids": [ 55 ], "parent_row_id": 55, "packet_id": 4, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 109.55194230769231, "packet_end": 138.7399278846154, "raw_text": "它们的核心思路", "selected_text": "它们的核心思路", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 55 ], "packet_id": 4, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 117.25194230769232, "end": 119.85194230769231, "text": "基本就是如何在做大模型的同时", "raw_asr_text": "基本就是如何在做大模型的同时", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 4, "words": [ { "word": "基本", "start": 117.25194230769232, "end": 117.75194230769232, "probability": 0.98291015625 }, { "word": "就是", "start": 117.75194230769232, "end": 118.07194230769231, "probability": 0.96875 }, { "word": "如何", "start": 118.07194230769231, "end": 118.43194230769231, "probability": 0.9892578125 }, { "word": "在", "start": 118.43194230769231, "end": 118.69194230769232, "probability": 0.98779296875 }, { "word": "做", "start": 118.69194230769232, "end": 118.91194230769231, "probability": 0.99365234375 }, { "word": "大", "start": 118.91194230769231, "end": 119.0919423076923, "probability": 0.99560546875 }, { "word": "模", "start": 119.0919423076923, "end": 119.27194230769231, "probability": 0.99853515625 }, { "word": "型", "start": 119.27194230769231, "end": 119.39194230769232, "probability": 0.9990234375 }, { "word": "的", "start": 119.39194230769232, "end": 119.49194230769231, "probability": 0.99658203125 }, { "word": "同", "start": 119.49194230769231, "end": 119.63194230769231, "probability": 0.998046875 }, { "word": "时", "start": 119.63194230769231, "end": 119.85194230769231, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.1259358913744267, "compression_ratio": 1.3447293447293447, "no_speech_prob": 5.938804309435497e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "高得多的竞争赛道因为做优化可比刷宝难多了在这款全新的V4.1 Flash模型中它们的核心思路基本就是如何在做大模型的同时大幅降低它的内存占用这是一个优化问题目标是降低大元模型的使用成本更具体的说是降低重度智能体场景的使用成本因为如今大元模型的大部分工作负载都需要反复读取塞满了工具输出和代码库的海量历史记录Deepsy基本是全新投入优化长上下" } ], "id": 56, "review_annotation": "", "row_id": 56, "bounded_row_id": 56, "parent_row_ids": [ 56 ], "parent_row_id": 56, "packet_id": 4, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 109.55194230769231, "packet_end": 138.7399278846154, "raw_text": "基本就是如何在做大模型的同时", "selected_text": "基本就是如何在做大模型的同时", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 56 ], "packet_id": 4, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 119.85194230769231, "end": 121.99194230769231, "text": "大幅降低它的内存占用", "raw_asr_text": "大幅降低它的内存占用", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 4, "words": [ { "word": "大", "start": 119.85194230769231, "end": 120.23194230769232, "probability": 0.99560546875 }, { "word": "幅", "start": 120.23194230769232, "end": 120.51194230769232, "probability": 1.0 }, { "word": "降", "start": 120.51194230769232, "end": 120.69194230769232, "probability": 0.9892578125 }, { "word": "低", "start": 120.69194230769232, "end": 121.01194230769232, "probability": 0.99853515625 }, { "word": "它的", "start": 121.01194230769232, "end": 121.25194230769232, "probability": 0.9794921875 }, { "word": "内", "start": 121.25194230769232, "end": 121.45194230769232, "probability": 0.99609375 }, { "word": "存", "start": 121.45194230769232, "end": 121.63194230769231, "probability": 0.99951171875 }, { "word": "占", "start": 121.63194230769231, "end": 121.79194230769231, "probability": 0.9921875 }, { "word": "用", "start": 121.79194230769231, "end": 121.99194230769231, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.1259358913744267, "compression_ratio": 1.3447293447293447, "no_speech_prob": 5.938804309435497e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "高得多的竞争赛道因为做优化可比刷宝难多了在这款全新的V4.1 Flash模型中它们的核心思路基本就是如何在做大模型的同时大幅降低它的内存占用这是一个优化问题目标是降低大元模型的使用成本更具体的说是降低重度智能体场景的使用成本因为如今大元模型的大部分工作负载都需要反复读取塞满了工具输出和代码库的海量历史记录Deepsy基本是全新投入优化长上下" } ], "id": 57, "review_annotation": "", "row_id": 57, "bounded_row_id": 57, "parent_row_ids": [ 57 ], "parent_row_id": 57, "packet_id": 4, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 109.55194230769231, "packet_end": 138.7399278846154, "raw_text": "大幅降低它的内存占用", "selected_text": "大幅降低它的内存占用", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 57 ], "packet_id": 4, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 121.99194230769231, "end": 123.27194230769231, "text": "这是一个优化问题", "raw_asr_text": "这是一个优化问题", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 4, "words": [ { "word": "这是", "start": 121.99194230769231, "end": 122.29194230769231, "probability": 0.990234375 }, { "word": "一个", "start": 122.29194230769231, "end": 122.55194230769231, "probability": 0.98779296875 }, { "word": "优", "start": 122.55194230769231, "end": 122.73194230769232, "probability": 0.9990234375 }, { "word": "化", "start": 122.73194230769232, "end": 122.93194230769231, "probability": 1.0 }, { "word": "问题", "start": 122.93194230769231, "end": 123.27194230769231, "probability": 0.9921875 } ], "asr_metrics": { "avg_logprob": -0.1259358913744267, "compression_ratio": 1.3447293447293447, "no_speech_prob": 5.938804309435497e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "高得多的竞争赛道因为做优化可比刷宝难多了在这款全新的V4.1 Flash模型中它们的核心思路基本就是如何在做大模型的同时大幅降低它的内存占用这是一个优化问题目标是降低大元模型的使用成本更具体的说是降低重度智能体场景的使用成本因为如今大元模型的大部分工作负载都需要反复读取塞满了工具输出和代码库的海量历史记录Deepsy基本是全新投入优化长上下" } ], "id": 58, "review_annotation": "", "row_id": 58, "bounded_row_id": 58, "parent_row_ids": [ 58 ], "parent_row_id": 58, "packet_id": 4, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 109.55194230769231, "packet_end": 138.7399278846154, "raw_text": "这是一个优化问题", "selected_text": "这是一个优化问题", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 58 ], "packet_id": 4, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 123.27194230769231, "end": 126.33194230769232, "text": "目标是降低大元模型的使用成本", "raw_asr_text": "目标是降低大元模型的使用成本", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 4, "words": [ { "word": "目", "start": 123.27194230769231, "end": 123.61194230769232, "probability": 0.99169921875 }, { "word": "标", "start": 123.61194230769232, "end": 123.87194230769231, "probability": 1.0 }, { "word": "是", "start": 123.87194230769231, "end": 124.25194230769232, "probability": 0.921875 }, { "word": "降", "start": 124.25194230769232, "end": 124.51194230769232, "probability": 0.9482421875 }, { "word": "低", "start": 124.51194230769232, "end": 124.79194230769231, "probability": 0.99951171875 }, { "word": "大", "start": 124.79194230769231, "end": 125.03194230769232, "probability": 0.99169921875 }, { "word": "元", "start": 125.03194230769232, "end": 125.23194230769232, "probability": 0.402587890625 }, { "word": "模", "start": 125.23194230769232, "end": 125.43194230769231, "probability": 0.99755859375 }, { "word": "型", "start": 125.43194230769231, "end": 125.5919423076923, "probability": 0.99951171875 }, { "word": "的", "start": 125.5919423076923, "end": 125.69194230769232, "probability": 0.9970703125 }, { "word": "使", "start": 125.69194230769232, "end": 125.85194230769231, "probability": 0.9990234375 }, { "word": "用", "start": 125.85194230769231, "end": 125.99194230769231, "probability": 1.0 }, { "word": "成", "start": 125.99194230769231, "end": 126.11194230769232, "probability": 0.99951171875 }, { "word": "本", "start": 126.11194230769232, "end": 126.33194230769232, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1259358913744267, "compression_ratio": 1.3447293447293447, "no_speech_prob": 5.938804309435497e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "高得多的竞争赛道因为做优化可比刷宝难多了在这款全新的V4.1 Flash模型中它们的核心思路基本就是如何在做大模型的同时大幅降低它的内存占用这是一个优化问题目标是降低大元模型的使用成本更具体的说是降低重度智能体场景的使用成本因为如今大元模型的大部分工作负载都需要反复读取塞满了工具输出和代码库的海量历史记录Deepsy基本是全新投入优化长上下" } ], "id": 59, "review_annotation": "", "row_id": 59, "bounded_row_id": 59, "parent_row_ids": [ 59 ], "parent_row_id": 59, "packet_id": 4, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 109.55194230769231, "packet_end": 138.7399278846154, "raw_text": "目标是降低大元模型的使用成本", "selected_text": "目标是降低大元模型的使用成本", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 59 ], "packet_id": 4, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 126.33194230769232, "end": 127.23194230769232, "text": "更具体的说", "raw_asr_text": "更具体的说", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 4, "words": [ { "word": "更", "start": 126.33194230769232, "end": 126.61194230769232, "probability": 0.99658203125 }, { "word": "具", "start": 126.61194230769232, "end": 126.81194230769232, "probability": 0.99951171875 }, { "word": "体", "start": 126.81194230769232, "end": 126.93194230769231, "probability": 0.99951171875 }, { "word": "的", "start": 126.93194230769231, "end": 127.03194230769232, "probability": 0.83935546875 }, { "word": "说", "start": 127.03194230769232, "end": 127.23194230769232, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.1259358913744267, "compression_ratio": 1.3447293447293447, "no_speech_prob": 5.938804309435497e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "高得多的竞争赛道因为做优化可比刷宝难多了在这款全新的V4.1 Flash模型中它们的核心思路基本就是如何在做大模型的同时大幅降低它的内存占用这是一个优化问题目标是降低大元模型的使用成本更具体的说是降低重度智能体场景的使用成本因为如今大元模型的大部分工作负载都需要反复读取塞满了工具输出和代码库的海量历史记录Deepsy基本是全新投入优化长上下" } ], "id": 60, "review_annotation": "", "row_id": 60, "bounded_row_id": 60, "parent_row_ids": [ 60 ], "parent_row_id": 60, "packet_id": 4, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 109.55194230769231, "packet_end": 138.7399278846154, "raw_text": "更具体的说", "selected_text": "更具体的说", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 60 ], "packet_id": 4, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 127.23194230769232, "end": 129.79194230769232, "text": "是降低重度智能体场景的使用成本", "raw_asr_text": "是降低重度智能体场景的使用成本", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 4, "words": [ { "word": "是", "start": 127.23194230769232, "end": 127.55194230769231, "probability": 0.98291015625 }, { "word": "降", "start": 127.55194230769231, "end": 127.79194230769231, "probability": 0.99462890625 }, { "word": "低", "start": 127.79194230769231, "end": 128.01194230769232, "probability": 0.99951171875 }, { "word": "重", "start": 128.01194230769232, "end": 128.2119423076923, "probability": 0.810546875 }, { "word": "度", "start": 128.2119423076923, "end": 128.3719423076923, "probability": 0.97900390625 }, { "word": "智", "start": 128.3719423076923, "end": 128.51194230769232, "probability": 0.998046875 }, { "word": "能", "start": 128.51194230769232, "end": 128.63194230769233, "probability": 0.99951171875 }, { "word": "体", "start": 128.63194230769233, "end": 128.7519423076923, "probability": 0.998046875 }, { "word": "场", "start": 128.7519423076923, "end": 128.9719423076923, "probability": 0.9912109375 }, { "word": "景", "start": 128.9719423076923, "end": 129.19194230769233, "probability": 1.0 }, { "word": "的", "start": 129.19194230769233, "end": 129.29194230769232, "probability": 0.99609375 }, { "word": "使", "start": 129.29194230769232, "end": 129.4119423076923, "probability": 0.99462890625 }, { "word": "用", "start": 129.4119423076923, "end": 129.51194230769232, "probability": 1.0 }, { "word": "成", "start": 129.51194230769232, "end": 129.63194230769233, "probability": 0.99951171875 }, { "word": "本", "start": 129.63194230769233, "end": 129.79194230769232, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1259358913744267, "compression_ratio": 1.3447293447293447, "no_speech_prob": 5.938804309435497e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "高得多的竞争赛道因为做优化可比刷宝难多了在这款全新的V4.1 Flash模型中它们的核心思路基本就是如何在做大模型的同时大幅降低它的内存占用这是一个优化问题目标是降低大元模型的使用成本更具体的说是降低重度智能体场景的使用成本因为如今大元模型的大部分工作负载都需要反复读取塞满了工具输出和代码库的海量历史记录Deepsy基本是全新投入优化长上下" } ], "id": 61, "review_annotation": "", "row_id": 61, "bounded_row_id": 61, "parent_row_ids": [ 61 ], "parent_row_id": 61, "packet_id": 4, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 109.55194230769231, "packet_end": 138.7399278846154, "raw_text": "是降低重度智能体场景的使用成本", "selected_text": "是降低重度智能体场景的使用成本", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 61 ], "packet_id": 4, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 129.79194230769232, "end": 132.23194230769232, "text": "因为如今大元模型的大部分工作负载", "raw_asr_text": "因为如今大元模型的大部分工作负载", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 4, "words": [ { "word": "因为", "start": 129.79194230769232, "end": 130.13194230769233, "probability": 0.99365234375 }, { "word": "如", "start": 130.13194230769233, "end": 130.35194230769233, "probability": 0.99560546875 }, { "word": "今", "start": 130.35194230769233, "end": 130.51194230769232, "probability": 0.99609375 }, { "word": "大", "start": 130.51194230769232, "end": 130.69194230769233, "probability": 0.9912109375 }, { "word": "元", "start": 130.69194230769233, "end": 130.8719423076923, "probability": 0.95361328125 }, { "word": "模", "start": 130.8719423076923, "end": 131.05194230769231, "probability": 0.9970703125 }, { "word": "型", "start": 131.05194230769231, "end": 131.17194230769232, "probability": 0.99951171875 }, { "word": "的", "start": 131.17194230769232, "end": 131.2719423076923, "probability": 0.9970703125 }, { "word": "大", "start": 131.2719423076923, "end": 131.4119423076923, "probability": 0.998046875 }, { "word": "部分", "start": 131.4119423076923, "end": 131.63194230769233, "probability": 0.99462890625 }, { "word": "工作", "start": 131.63194230769233, "end": 131.89194230769232, "probability": 0.9892578125 }, { "word": "负", "start": 131.89194230769232, "end": 132.05194230769231, "probability": 0.97998046875 }, { "word": "载", "start": 132.05194230769231, "end": 132.23194230769232, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.1259358913744267, "compression_ratio": 1.3447293447293447, "no_speech_prob": 5.938804309435497e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "高得多的竞争赛道因为做优化可比刷宝难多了在这款全新的V4.1 Flash模型中它们的核心思路基本就是如何在做大模型的同时大幅降低它的内存占用这是一个优化问题目标是降低大元模型的使用成本更具体的说是降低重度智能体场景的使用成本因为如今大元模型的大部分工作负载都需要反复读取塞满了工具输出和代码库的海量历史记录Deepsy基本是全新投入优化长上下" } ], "id": 62, "review_annotation": "", "row_id": 62, "bounded_row_id": 62, "parent_row_ids": [ 62 ], "parent_row_id": 62, "packet_id": 4, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 109.55194230769231, "packet_end": 138.7399278846154, "raw_text": "因为如今大元模型的大部分工作负载", "selected_text": "因为如今大元模型的大部分工作负载", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 62 ], "packet_id": 4, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 132.23194230769232, "end": 133.2119423076923, "text": "都需要反复读取", "raw_asr_text": "都需要反复读取", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 4, "words": [ { "word": "都", "start": 132.23194230769232, "end": 132.45194230769232, "probability": 0.99462890625 }, { "word": "需要", "start": 132.45194230769232, "end": 132.63194230769233, "probability": 0.99560546875 }, { "word": "反", "start": 132.63194230769233, "end": 132.7719423076923, "probability": 0.99609375 }, { "word": "复", "start": 132.7719423076923, "end": 132.9319423076923, "probability": 0.998046875 }, { "word": "读", "start": 132.9319423076923, "end": 133.0319423076923, "probability": 0.99853515625 }, { "word": "取", "start": 133.0319423076923, "end": 133.2119423076923, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1259358913744267, "compression_ratio": 1.3447293447293447, "no_speech_prob": 5.938804309435497e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "高得多的竞争赛道因为做优化可比刷宝难多了在这款全新的V4.1 Flash模型中它们的核心思路基本就是如何在做大模型的同时大幅降低它的内存占用这是一个优化问题目标是降低大元模型的使用成本更具体的说是降低重度智能体场景的使用成本因为如今大元模型的大部分工作负载都需要反复读取塞满了工具输出和代码库的海量历史记录Deepsy基本是全新投入优化长上下" } ], "id": 63, "review_annotation": "", "row_id": 63, "bounded_row_id": 63, "parent_row_ids": [ 63 ], "parent_row_id": 63, "packet_id": 4, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 109.55194230769231, "packet_end": 138.7399278846154, "raw_text": "都需要反复读取", "selected_text": "都需要反复读取", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 63 ], "packet_id": 4, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 133.2119423076923, "end": 135.7519423076923, "text": "塞满了工具输出和代码库的海量历史记录", "raw_asr_text": "塞满了工具输出和代码库的海量历史记录", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 4, "words": [ { "word": "塞", "start": 133.2119423076923, "end": 133.51194230769232, "probability": 0.98876953125 }, { "word": "满", "start": 133.51194230769232, "end": 133.63194230769233, "probability": 0.99755859375 }, { "word": "了", "start": 133.63194230769233, "end": 133.73194230769232, "probability": 0.9326171875 }, { "word": "工", "start": 133.73194230769232, "end": 133.89194230769232, "probability": 0.99658203125 }, { "word": "具", "start": 133.89194230769232, "end": 133.9919423076923, "probability": 0.9990234375 }, { "word": "输", "start": 133.9919423076923, "end": 134.17194230769232, "probability": 0.9931640625 }, { "word": "出", "start": 134.17194230769232, "end": 134.3719423076923, "probability": 0.998046875 }, { "word": "和", "start": 134.3719423076923, "end": 134.57194230769232, "probability": 0.994140625 }, { "word": "代", "start": 134.57194230769232, "end": 134.67194230769232, "probability": 0.998046875 }, { "word": "码", "start": 134.67194230769232, "end": 134.8119423076923, "probability": 0.99609375 }, { "word": "库", "start": 134.8119423076923, "end": 134.9119423076923, "probability": 0.99609375 }, { "word": "的", "start": 134.9119423076923, "end": 135.0319423076923, "probability": 0.998046875 }, { "word": "海", "start": 135.0319423076923, "end": 135.19194230769233, "probability": 0.9951171875 }, { "word": "量", "start": 135.19194230769233, "end": 135.33194230769232, "probability": 0.99169921875 }, { "word": "历", "start": 135.33194230769232, "end": 135.45194230769232, "probability": 0.99609375 }, { "word": "史", "start": 135.45194230769232, "end": 135.55194230769231, "probability": 0.99951171875 }, { "word": "记", "start": 135.55194230769231, "end": 135.6519423076923, "probability": 0.96728515625 }, { "word": "录", "start": 135.6519423076923, "end": 135.7519423076923, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1259358913744267, "compression_ratio": 1.3447293447293447, "no_speech_prob": 5.938804309435497e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "高得多的竞争赛道因为做优化可比刷宝难多了在这款全新的V4.1 Flash模型中它们的核心思路基本就是如何在做大模型的同时大幅降低它的内存占用这是一个优化问题目标是降低大元模型的使用成本更具体的说是降低重度智能体场景的使用成本因为如今大元模型的大部分工作负载都需要反复读取塞满了工具输出和代码库的海量历史记录Deepsy基本是全新投入优化长上下" } ], "id": 64, "review_annotation": "", "row_id": 64, "bounded_row_id": 64, "parent_row_ids": [ 64 ], "parent_row_id": 64, "packet_id": 4, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 109.55194230769231, "packet_end": 138.7399278846154, "raw_text": "塞满了工具输出和代码库的海量历史记录", "selected_text": "塞满了工具输出和代码库的海量历史记录", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 64 ], "packet_id": 4, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 135.7519423076923, "end": 137.9399278846154, "text": "Deepsy基本是全新投入优化", "raw_asr_text": "Deepsy基本是全新投入优化", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 4, "words": [ { "word": "De", "start": 135.7519423076923, "end": 135.85194230769233, "probability": 0.701171875 }, { "word": "ep", "start": 135.85194230769233, "end": 135.95194230769232, "probability": 0.95703125 }, { "word": "sy", "start": 135.95194230769232, "end": 136.0919423076923, "probability": 0.4609375 }, { "word": "基本", "start": 136.0919423076923, "end": 136.3719423076923, "probability": 0.974609375 }, { "word": "是", "start": 136.3719423076923, "end": 136.63194230769233, "probability": 0.99365234375 }, { "word": "全", "start": 136.63194230769233, "end": 136.95194230769232, "probability": 0.986328125 }, { "word": "新", "start": 136.95194230769232, "end": 137.17194230769232, "probability": 0.994140625 }, { "word": "投", "start": 137.17194230769232, "end": 137.39194230769232, "probability": 0.9931640625 }, { "word": "入", "start": 137.39194230769232, "end": 137.5919423076923, "probability": 0.99951171875 }, { "word": "优", "start": 137.5919423076923, "end": 137.8119423076923, "probability": 0.998046875 }, { "word": "化", "start": 137.8119423076923, "end": 138.0319423076923, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1259358913744267, "compression_ratio": 1.3447293447293447, "no_speech_prob": 5.938804309435497e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "高得多的竞争赛道因为做优化可比刷宝难多了在这款全新的V4.1 Flash模型中它们的核心思路基本就是如何在做大模型的同时大幅降低它的内存占用这是一个优化问题目标是降低大元模型的使用成本更具体的说是降低重度智能体场景的使用成本因为如今大元模型的大部分工作负载都需要反复读取塞满了工具输出和代码库的海量历史记录Deepsy基本是全新投入优化长上下" } ], "id": 65, "review_annotation": "", "row_id": 65, "bounded_row_id": 65, "parent_row_ids": [ 65 ], "parent_row_id": 65, "packet_id": 4, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 109.55194230769231, "packet_end": 138.7399278846154, "raw_text": "Deepsy基本是全新投入优化", "selected_text": "Deepsy基本是全新投入优化", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 65 ], "packet_id": 4, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 137.9399278846154, "end": 139.1599278846154, "text": "投入优化 常上下文下", "raw_asr_text": "投入优化 常上下文下", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "投", "start": 137.1399278846154, "end": 137.3799278846154, "probability": 0.421875 }, { "word": "入", "start": 137.3799278846154, "end": 137.5799278846154, "probability": 0.9990234375 }, { "word": "优", "start": 137.5799278846154, "end": 137.77992788461538, "probability": 0.98828125 }, { "word": "化", "start": 137.77992788461538, "end": 138.01992788461538, "probability": 0.99951171875 }, { "word": " 常", "start": 138.2599278846154, "end": 138.39992788461538, "probability": 0.5380859375 }, { "word": "上", "start": 138.39992788461538, "end": 138.61992788461538, "probability": 0.99658203125 }, { "word": "下", "start": 138.61992788461538, "end": 138.77992788461538, "probability": 0.9931640625 }, { "word": "文", "start": 138.77992788461538, "end": 138.95992788461538, "probability": 0.87060546875 }, { "word": "下", "start": 138.95992788461538, "end": 139.1599278846154, "probability": 0.95947265625 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 66, "review_annotation": "", "row_id": 66, "bounded_row_id": 66, "parent_row_ids": [ 66 ], "parent_row_id": 66, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "投入优化 常上下文下", "selected_text": "投入优化 常上下文下", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 66 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 139.1599278846154, "end": 140.7199278846154, "text": "注意力开销过大的问题", "raw_asr_text": "注意力开销过大的问题", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "注意", "start": 139.1599278846154, "end": 139.4999278846154, "probability": 0.54638671875 }, { "word": "力", "start": 139.4999278846154, "end": 139.6599278846154, "probability": 0.998046875 }, { "word": "开", "start": 139.6599278846154, "end": 139.83992788461538, "probability": 0.9951171875 }, { "word": "销", "start": 139.83992788461538, "end": 140.05992788461538, "probability": 0.9931640625 }, { "word": "过", "start": 140.05992788461538, "end": 140.1999278846154, "probability": 0.994140625 }, { "word": "大的", "start": 140.1999278846154, "end": 140.39992788461538, "probability": 0.9619140625 }, { "word": "问题", "start": 140.39992788461538, "end": 140.7199278846154, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 67, "review_annotation": "", "row_id": 67, "bounded_row_id": 67, "parent_row_ids": [ 67 ], "parent_row_id": 67, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "注意力开销过大的问题", "selected_text": "注意力开销过大的问题", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 67 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 140.7199278846154, "end": 142.73992788461538, "text": "虽然说他们现有的稀疏注意力方案", "raw_asr_text": "虽然说他们现有的稀疏注意力方案", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "虽", "start": 140.7199278846154, "end": 141.0799278846154, "probability": 0.7158203125 }, { "word": "然", "start": 141.0799278846154, "end": 141.23992788461538, "probability": 0.417724609375 }, { "word": "说", "start": 141.23992788461538, "end": 141.2599278846154, "probability": 0.5576171875 }, { "word": "他们", "start": 141.2599278846154, "end": 141.4999278846154, "probability": 0.78369140625 }, { "word": "现", "start": 141.4999278846154, "end": 141.6399278846154, "probability": 0.9853515625 }, { "word": "有", "start": 141.6399278846154, "end": 141.79992788461539, "probability": 0.9990234375 }, { "word": "的", "start": 141.79992788461539, "end": 141.89992788461538, "probability": 0.994140625 }, { "word": "稀", "start": 141.89992788461538, "end": 141.9799278846154, "probability": 0.6689453125 }, { "word": "疏", "start": 141.9799278846154, "end": 142.1399278846154, "probability": 0.97265625 }, { "word": "注意", "start": 142.1399278846154, "end": 142.3199278846154, "probability": 0.9736328125 }, { "word": "力", "start": 142.3199278846154, "end": 142.4199278846154, "probability": 0.9990234375 }, { "word": "方", "start": 142.4199278846154, "end": 142.55992788461538, "probability": 0.978515625 }, { "word": "案", "start": 142.55992788461538, "end": 142.73992788461538, "probability": 0.9921875 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 68, "review_annotation": "", "row_id": 68, "bounded_row_id": 68, "parent_row_ids": [ 68 ], "parent_row_id": 68, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "虽然说他们现有的稀疏注意力方案", "selected_text": "虽然说他们现有的稀疏注意力方案", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 68 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 142.73992788461538, "end": 144.2599278846154, "text": "已经降低了注意力成本", "raw_asr_text": "已经降低了注意力成本", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "已经", "start": 142.73992788461538, "end": 143.0399278846154, "probability": 0.97265625 }, { "word": "降", "start": 143.0399278846154, "end": 143.2599278846154, "probability": 0.9951171875 }, { "word": "低", "start": 143.2599278846154, "end": 143.4999278846154, "probability": 0.990234375 }, { "word": "了", "start": 143.4999278846154, "end": 143.5999278846154, "probability": 0.9931640625 }, { "word": "注意", "start": 143.5999278846154, "end": 143.7599278846154, "probability": 0.99169921875 }, { "word": "力", "start": 143.7599278846154, "end": 143.9199278846154, "probability": 0.99853515625 }, { "word": "成", "start": 143.9199278846154, "end": 144.05992788461538, "probability": 0.99658203125 }, { "word": "本", "start": 144.05992788461538, "end": 144.2599278846154, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 69, "review_annotation": "", "row_id": 69, "bounded_row_id": 69, "parent_row_ids": [ 69 ], "parent_row_id": 69, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "已经降低了注意力成本", "selected_text": "已经降低了注意力成本", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 69 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 144.2599278846154, "end": 145.8199278846154, "text": "但另一个问题开始浮现", "raw_asr_text": "但另一个问题开始浮现", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "但", "start": 144.2599278846154, "end": 144.55992788461538, "probability": 0.95166015625 }, { "word": "另", "start": 144.55992788461538, "end": 144.6999278846154, "probability": 0.990234375 }, { "word": "一个", "start": 144.6999278846154, "end": 144.8599278846154, "probability": 0.994140625 }, { "word": "问题", "start": 144.8599278846154, "end": 145.11992788461538, "probability": 0.99853515625 }, { "word": "开始", "start": 145.11992788461538, "end": 145.4199278846154, "probability": 0.9912109375 }, { "word": "浮", "start": 145.4199278846154, "end": 145.61992788461538, "probability": 0.9990234375 }, { "word": "现", "start": 145.61992788461538, "end": 145.8199278846154, "probability": 0.9970703125 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 70, "review_annotation": "", "row_id": 70, "bounded_row_id": 70, "parent_row_ids": [ 70 ], "parent_row_id": 70, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "但另一个问题开始浮现", "selected_text": "但另一个问题开始浮现", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 70 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 145.8199278846154, "end": 147.9399278846154, "text": "那就是模型仍然需要存储", "raw_asr_text": "那就是模型仍然需要存储", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "那", "start": 145.8199278846154, "end": 146.29992788461539, "probability": 0.88525390625 }, { "word": "就是", "start": 146.29992788461539, "end": 146.49992788461537, "probability": 0.99267578125 }, { "word": "模", "start": 146.49992788461537, "end": 146.6999278846154, "probability": 0.9951171875 }, { "word": "型", "start": 146.6999278846154, "end": 146.89992788461538, "probability": 0.9990234375 }, { "word": "仍", "start": 146.89992788461538, "end": 147.0799278846154, "probability": 0.998046875 }, { "word": "然", "start": 147.0799278846154, "end": 147.17992788461538, "probability": 0.994140625 }, { "word": "需要", "start": 147.17992788461538, "end": 147.3599278846154, "probability": 0.9921875 }, { "word": "存", "start": 147.3599278846154, "end": 147.6599278846154, "probability": 0.99755859375 }, { "word": "储", "start": 147.6599278846154, "end": 147.9399278846154, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 71, "review_annotation": "", "row_id": 71, "bounded_row_id": 71, "parent_row_ids": [ 71 ], "parent_row_id": 71, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "那就是模型仍然需要存储", "selected_text": "那就是模型仍然需要存储", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 71 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 147.9399278846154, "end": 150.8199278846154, "text": "搬运重新加载巨量的KV缓存", "raw_asr_text": "搬运重新加载巨量的KV缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "搬", "start": 147.9399278846154, "end": 148.29992788461539, "probability": 0.9716796875 }, { "word": "运", "start": 148.29992788461539, "end": 148.5799278846154, "probability": 0.9970703125 }, { "word": "重", "start": 148.5799278846154, "end": 148.8199278846154, "probability": 0.8603515625 }, { "word": "新", "start": 148.8199278846154, "end": 149.11992788461538, "probability": 0.9990234375 }, { "word": "加", "start": 149.11992788461538, "end": 149.33992788461538, "probability": 0.98583984375 }, { "word": "载", "start": 149.33992788461538, "end": 149.55992788461538, "probability": 0.998046875 }, { "word": "巨", "start": 149.55992788461538, "end": 149.89992788461538, "probability": 0.9951171875 }, { "word": "量", "start": 149.89992788461538, "end": 150.0799278846154, "probability": 0.99755859375 }, { "word": "的", "start": 150.0799278846154, "end": 150.1999278846154, "probability": 0.9814453125 }, { "word": "K", "start": 150.1999278846154, "end": 150.33992788461538, "probability": 0.482666015625 }, { "word": "V", "start": 150.33992788461538, "end": 150.49992788461537, "probability": 0.98583984375 }, { "word": "缓", "start": 150.49992788461537, "end": 150.6399278846154, "probability": 0.994140625 }, { "word": "存", "start": 150.6399278846154, "end": 150.8199278846154, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 72, "review_annotation": "", "row_id": 72, "bounded_row_id": 72, "parent_row_ids": [ 72 ], "parent_row_id": 72, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "搬运重新加载巨量的KV缓存", "selected_text": "搬运重新加载巨量的KV缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 72 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 150.8199278846154, "end": 152.01992788461538, "text": "下面我快速讲下", "raw_asr_text": "下面我快速讲下", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "下面", "start": 150.8199278846154, "end": 151.17992788461538, "probability": 0.951171875 }, { "word": "我", "start": 151.17992788461538, "end": 151.3799278846154, "probability": 0.9853515625 }, { "word": "快", "start": 151.3799278846154, "end": 151.51992788461538, "probability": 0.99658203125 }, { "word": "速", "start": 151.51992788461538, "end": 151.6599278846154, "probability": 0.99853515625 }, { "word": "讲", "start": 151.6599278846154, "end": 151.8199278846154, "probability": 0.99658203125 }, { "word": "下", "start": 151.8199278846154, "end": 152.01992788461538, "probability": 0.82470703125 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 73, "review_annotation": "", "row_id": 73, "bounded_row_id": 73, "parent_row_ids": [ 73 ], "parent_row_id": 73, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "下面我快速讲下", "selected_text": "下面我快速讲下", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 73 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 152.01992788461538, "end": 153.83992788461538, "text": "KV缓存和注意力的关系", "raw_asr_text": "KV缓存和注意力的关系", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "K", "start": 152.01992788461538, "end": 152.29992788461539, "probability": 0.814453125 }, { "word": "V", "start": 152.29992788461539, "end": 152.4399278846154, "probability": 0.9990234375 }, { "word": "缓", "start": 152.4399278846154, "end": 152.6399278846154, "probability": 0.998046875 }, { "word": "存", "start": 152.6399278846154, "end": 152.79992788461539, "probability": 1.0 }, { "word": "和", "start": 152.79992788461539, "end": 152.9799278846154, "probability": 0.98583984375 }, { "word": "注意", "start": 152.9799278846154, "end": 153.17992788461538, "probability": 0.9970703125 }, { "word": "力", "start": 153.17992788461538, "end": 153.3199278846154, "probability": 0.99755859375 }, { "word": "的", "start": 153.3199278846154, "end": 153.4199278846154, "probability": 0.986328125 }, { "word": "关", "start": 153.4199278846154, "end": 153.5799278846154, "probability": 0.9990234375 }, { "word": "系", "start": 153.5799278846154, "end": 153.83992788461538, "probability": 0.9970703125 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 74, "review_annotation": "", "row_id": 74, "bounded_row_id": 74, "parent_row_ids": [ 74 ], "parent_row_id": 74, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "KV缓存和注意力的关系", "selected_text": "KV缓存和注意力的关系", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 74 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 153.83992788461538, "end": 155.5999278846154, "text": "大元模型生成一个Token时", "raw_asr_text": "大元模型生成一个Token时", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "大", "start": 153.83992788461538, "end": 154.0799278846154, "probability": 0.96240234375 }, { "word": "元", "start": 154.0799278846154, "end": 154.23992788461538, "probability": 0.403076171875 }, { "word": "模", "start": 154.23992788461538, "end": 154.39992788461538, "probability": 0.9990234375 }, { "word": "型", "start": 154.39992788461538, "end": 154.55992788461538, "probability": 0.99755859375 }, { "word": "生", "start": 154.55992788461538, "end": 154.73992788461538, "probability": 0.9462890625 }, { "word": "成", "start": 154.73992788461538, "end": 154.89992788461538, "probability": 0.998046875 }, { "word": "一个", "start": 154.89992788461538, "end": 155.0999278846154, "probability": 0.98779296875 }, { "word": "T", "start": 155.0999278846154, "end": 155.2599278846154, "probability": 0.54638671875 }, { "word": "oken", "start": 155.2599278846154, "end": 155.4199278846154, "probability": 0.91015625 }, { "word": "时", "start": 155.4199278846154, "end": 155.5999278846154, "probability": 0.70849609375 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 75, "review_annotation": "", "row_id": 75, "bounded_row_id": 75, "parent_row_ids": [ 75 ], "parent_row_id": 75, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "大元模型生成一个Token时", "selected_text": "大元模型生成一个Token时", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 75 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 155.5999278846154, "end": 157.2199278846154, "text": "每个注意力层都需要用到", "raw_asr_text": "每个注意力层都需要用到", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "每", "start": 155.5999278846154, "end": 155.89992788461538, "probability": 0.958984375 }, { "word": "个", "start": 155.89992788461538, "end": 156.01992788461538, "probability": 0.99658203125 }, { "word": "注意", "start": 156.01992788461538, "end": 156.1999278846154, "probability": 0.98583984375 }, { "word": "力", "start": 156.1999278846154, "end": 156.33992788461538, "probability": 0.99609375 }, { "word": "层", "start": 156.33992788461538, "end": 156.51992788461538, "probability": 0.9970703125 }, { "word": "都", "start": 156.51992788461538, "end": 156.67992788461538, "probability": 0.994140625 }, { "word": "需要", "start": 156.67992788461538, "end": 156.89992788461538, "probability": 0.99462890625 }, { "word": "用", "start": 156.89992788461538, "end": 157.05992788461538, "probability": 0.99267578125 }, { "word": "到", "start": 157.05992788461538, "end": 157.2199278846154, "probability": 0.99365234375 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 76, "review_annotation": "", "row_id": 76, "bounded_row_id": 76, "parent_row_ids": [ 76 ], "parent_row_id": 76, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "每个注意力层都需要用到", "selected_text": "每个注意力层都需要用到", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 76 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 157.2199278846154, "end": 159.2799278846154, "text": "此前所有Token的缓存键和纸", "raw_asr_text": "此前所有Token的缓存键和纸", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "此", "start": 157.2199278846154, "end": 157.4799278846154, "probability": 0.998046875 }, { "word": "前", "start": 157.4799278846154, "end": 157.67992788461538, "probability": 0.99951171875 }, { "word": "所有", "start": 157.67992788461538, "end": 157.9799278846154, "probability": 0.99365234375 }, { "word": "T", "start": 157.9799278846154, "end": 158.17992788461538, "probability": 0.630859375 }, { "word": "oken", "start": 158.17992788461538, "end": 158.33992788461538, "probability": 0.8173828125 }, { "word": "的", "start": 158.33992788461538, "end": 158.51992788461538, "probability": 0.67578125 }, { "word": "缓", "start": 158.51992788461538, "end": 158.67992788461538, "probability": 0.9970703125 }, { "word": "存", "start": 158.67992788461538, "end": 158.79992788461539, "probability": 0.99853515625 }, { "word": "键", "start": 158.79992788461539, "end": 158.95992788461538, "probability": 0.8505859375 }, { "word": "和", "start": 158.95992788461538, "end": 159.0999278846154, "probability": 0.9912109375 }, { "word": "纸", "start": 159.0999278846154, "end": 159.2799278846154, "probability": 0.9658203125 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 77, "review_annotation": "", "row_id": 77, "bounded_row_id": 77, "parent_row_ids": [ 77 ], "parent_row_id": 77, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "此前所有Token的缓存键和纸", "selected_text": "此前所有Token的缓存键和纸", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 77 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 159.2799278846154, "end": 160.67992788461538, "text": "键主要用来帮它判断", "raw_asr_text": "键主要用来帮它判断", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "键", "start": 159.2799278846154, "end": 159.55992788461538, "probability": 0.97119140625 }, { "word": "主", "start": 159.55992788461538, "end": 159.73992788461538, "probability": 0.94140625 }, { "word": "要", "start": 159.73992788461538, "end": 159.83992788461538, "probability": 0.99951171875 }, { "word": "用", "start": 159.83992788461538, "end": 159.9399278846154, "probability": 0.98486328125 }, { "word": "来", "start": 159.9399278846154, "end": 160.05992788461538, "probability": 0.994140625 }, { "word": "帮", "start": 160.05992788461538, "end": 160.1999278846154, "probability": 0.994140625 }, { "word": "它", "start": 160.1999278846154, "end": 160.3599278846154, "probability": 0.603515625 }, { "word": "判", "start": 160.3599278846154, "end": 160.51992788461538, "probability": 0.99853515625 }, { "word": "断", "start": 160.51992788461538, "end": 160.67992788461538, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 78, "review_annotation": "", "row_id": 78, "bounded_row_id": 78, "parent_row_ids": [ 78 ], "parent_row_id": 78, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "键主要用来帮它判断", "selected_text": "键主要用来帮它判断", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 78 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 160.67992788461538, "end": 162.33992788461538, "text": "哪些历史信息是相关的", "raw_asr_text": "哪些历史信息是相关的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "哪", "start": 160.67992788461538, "end": 160.89992788461538, "probability": 0.98876953125 }, { "word": "些", "start": 160.89992788461538, "end": 161.11992788461538, "probability": 0.99853515625 }, { "word": "历", "start": 161.11992788461538, "end": 161.2799278846154, "probability": 0.9990234375 }, { "word": "史", "start": 161.2799278846154, "end": 161.39992788461538, "probability": 0.99951171875 }, { "word": "信", "start": 161.39992788461538, "end": 161.51992788461538, "probability": 0.9951171875 }, { "word": "息", "start": 161.51992788461538, "end": 161.6599278846154, "probability": 1.0 }, { "word": "是", "start": 161.6599278846154, "end": 161.79992788461539, "probability": 0.99609375 }, { "word": "相", "start": 161.79992788461539, "end": 161.95992788461538, "probability": 0.99951171875 }, { "word": "关", "start": 161.95992788461538, "end": 162.1599278846154, "probability": 0.99951171875 }, { "word": "的", "start": 162.1599278846154, "end": 162.33992788461538, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 79, "review_annotation": "", "row_id": 79, "bounded_row_id": 79, "parent_row_ids": [ 79 ], "parent_row_id": 79, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "哪些历史信息是相关的", "selected_text": "哪些历史信息是相关的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 79 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 162.33992788461538, "end": 164.6599278846154, "text": "职责包含它实际要调取的信息", "raw_asr_text": "职责包含它实际要调取的信息", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "职", "start": 162.33992788461538, "end": 162.6399278846154, "probability": 0.7890625 }, { "word": "责", "start": 162.6399278846154, "end": 162.8599278846154, "probability": 0.9482421875 }, { "word": "包", "start": 162.8599278846154, "end": 163.01992788461538, "probability": 0.99658203125 }, { "word": "含", "start": 163.01992788461538, "end": 163.1599278846154, "probability": 0.9970703125 }, { "word": "它", "start": 163.1599278846154, "end": 163.3199278846154, "probability": 0.9755859375 }, { "word": "实", "start": 163.3199278846154, "end": 163.55992788461538, "probability": 0.99072265625 }, { "word": "际", "start": 163.55992788461538, "end": 163.6999278846154, "probability": 0.9990234375 }, { "word": "要", "start": 163.6999278846154, "end": 163.8599278846154, "probability": 0.98193359375 }, { "word": "调", "start": 163.8599278846154, "end": 164.0799278846154, "probability": 0.986328125 }, { "word": "取", "start": 164.0799278846154, "end": 164.2199278846154, "probability": 0.99951171875 }, { "word": "的", "start": 164.2199278846154, "end": 164.33992788461538, "probability": 0.9970703125 }, { "word": "信", "start": 164.33992788461538, "end": 164.45992788461538, "probability": 0.99609375 }, { "word": "息", "start": 164.45992788461538, "end": 164.6599278846154, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 80, "review_annotation": "", "row_id": 80, "bounded_row_id": 80, "parent_row_ids": [ 80 ], "parent_row_id": 80, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "职责包含它实际要调取的信息", "selected_text": "职责包含它实际要调取的信息", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 80 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 164.6599278846154, "end": 165.52791346153847, "text": "但要是每生成一个Token", "raw_asr_text": "但要是每生成一个Token", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 5, "words": [ { "word": "但", "start": 164.6599278846154, "end": 165.05992788461538, "probability": 0.9765625 }, { "word": "要", "start": 165.05992788461538, "end": 165.2199278846154, "probability": 0.994140625 }, { "word": "是", "start": 165.2199278846154, "end": 165.39992788461538, "probability": 0.99609375 }, { "word": "每", "start": 165.39992788461538, "end": 165.5399278846154, "probability": 0.998046875 }, { "word": "生", "start": 165.5399278846154, "end": 165.7199278846154, "probability": 0.98388671875 }, { "word": "成", "start": 165.7199278846154, "end": 165.8799278846154, "probability": 0.98876953125 }, { "word": "一个", "start": 165.8799278846154, "end": 166.0999278846154, "probability": 0.98583984375 }, { "word": "T", "start": 166.0999278846154, "end": 166.23992788461538, "probability": 0.95947265625 }, { "word": "oken", "start": 166.23992788461538, "end": 166.29992788461539, "probability": 0.984375 } ], "asr_metrics": { "avg_logprob": -0.11939136187235515, "compression_ratio": 1.3977900552486189, "no_speech_prob": 5.514496845271388e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投入优化 常上下文下注意力开销过大的问题虽然说他们现有的稀疏注意力方案已经降低了注意力成本但另一个问题开始浮现那就是模型仍然需要存储搬运重新加载巨量的KV缓存下面我快速讲下KV缓存和注意力的关系大元模型生成一个Token时每个注意力层都需要用到此前所有Token的缓存键和纸键主要用来帮它判断哪些历史信息是相关的职责包含它实际要调取的信息但要是每生成一个Token" } ], "id": 81, "review_annotation": "", "row_id": 81, "bounded_row_id": 81, "parent_row_ids": [ 81 ], "parent_row_id": 81, "packet_id": 5, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 137.1399278846154, "packet_end": 166.3279134615385, "raw_text": "但要是每生成一个Token", "selected_text": "但要是每生成一个Token", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 81 ], "packet_id": 5, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 165.52791346153847, "end": 169.56791346153847, "text": "但要是每生成一个token都重新计算所有这些表征浪费会大的离谱", "raw_asr_text": "但要是每生成一个token都重新计算所有这些表征浪费会大的离谱", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 6, "words": [ { "word": "但", "start": 164.72791346153846, "end": 165.06791346153847, "probability": 0.97265625 }, { "word": "要", "start": 165.06791346153847, "end": 165.22791346153846, "probability": 0.9892578125 }, { "word": "是", "start": 165.22791346153846, "end": 165.38791346153846, "probability": 0.99169921875 }, { "word": "每", "start": 165.38791346153846, "end": 165.54791346153846, "probability": 0.99658203125 }, { "word": "生", "start": 165.54791346153846, "end": 165.70791346153845, "probability": 0.974609375 }, { "word": "成", "start": 165.70791346153845, "end": 165.86791346153845, "probability": 0.857421875 }, { "word": "一个", "start": 165.86791346153845, "end": 166.06791346153847, "probability": 0.859375 }, { "word": "t", "start": 166.06791346153847, "end": 166.26791346153846, "probability": 0.52734375 }, { "word": "oken", "start": 166.26791346153846, "end": 166.46791346153847, "probability": 0.99462890625 }, { "word": "都", "start": 166.46791346153847, "end": 166.68791346153847, "probability": 0.9287109375 }, { "word": "重", "start": 166.68791346153847, "end": 166.92791346153845, "probability": 0.9951171875 }, { "word": "新", "start": 166.92791346153845, "end": 167.10791346153846, "probability": 0.99853515625 }, { "word": "计", "start": 167.10791346153846, "end": 167.26791346153846, "probability": 0.99609375 }, { "word": "算", "start": 167.26791346153846, "end": 167.40791346153847, "probability": 1.0 }, { "word": "所有", "start": 167.40791346153847, "end": 167.66791346153846, "probability": 0.96728515625 }, { "word": "这", "start": 167.66791346153846, "end": 167.82791346153846, "probability": 0.9541015625 }, { "word": "些", "start": 167.82791346153846, "end": 167.94791346153846, "probability": 0.99951171875 }, { "word": "表", "start": 167.94791346153846, "end": 168.10791346153846, "probability": 0.9970703125 }, { "word": "征", "start": 168.10791346153846, "end": 168.34791346153847, "probability": 0.859375 }, { "word": "浪", "start": 168.34791346153847, "end": 168.60791346153846, "probability": 0.98046875 }, { "word": "费", "start": 168.60791346153846, "end": 168.74791346153847, "probability": 1.0 }, { "word": "会", "start": 168.74791346153847, "end": 168.90791346153847, "probability": 0.99609375 }, { "word": "大的", "start": 168.90791346153847, "end": 169.12791346153847, "probability": 0.8642578125 }, { "word": "离", "start": 169.12791346153847, "end": 169.38791346153846, "probability": 0.99755859375 }, { "word": "谱", "start": 169.38791346153846, "end": 169.56791346153847, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.084868532940022, "compression_ratio": 1.3660714285714286, "no_speech_prob": 2.4770080084279833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但要是每生成一个token都重新计算所有这些表征浪费会大的离谱所以模型会把它们存下来这些存储下来的状态就是kb缓存token越多kb缓存就越大层数越多kb缓存也越大所以当上下文规模达到100万token时瓶颈就不再只是算力速度了kb缓存带来的延迟和成本问题会逐渐凸显你需要一次性计算搬运的token数量是注意力机制最初设计承载量的10到100倍所有欲填充" } ], "id": 82, "review_annotation": "", "row_id": 82, "bounded_row_id": 82, "parent_row_ids": [ 82 ], "parent_row_id": 82, "packet_id": 6, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 164.72791346153846, "packet_end": 193.91589903846156, "raw_text": "但要是每生成一个token都重新计算所有这些表征浪费会大的离谱", "selected_text": "但要是每生成一个token都重新计算所有这些表征浪费会大的离谱", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 82 ], "packet_id": 6, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 169.56791346153847, "end": 171.20791346153845, "text": "所以模型会把它们存下来", "raw_asr_text": "所以模型会把它们存下来", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 6, "words": [ { "word": "所以", "start": 169.56791346153847, "end": 169.84791346153847, "probability": 0.9306640625 }, { "word": "模", "start": 169.84791346153847, "end": 170.04791346153846, "probability": 0.998046875 }, { "word": "型", "start": 170.04791346153846, "end": 170.20791346153845, "probability": 0.99951171875 }, { "word": "会", "start": 170.20791346153845, "end": 170.34791346153847, "probability": 0.99853515625 }, { "word": "把它", "start": 170.34791346153847, "end": 170.46791346153847, "probability": 0.9521484375 }, { "word": "们", "start": 170.46791346153847, "end": 170.66791346153846, "probability": 0.99853515625 }, { "word": "存", "start": 170.66791346153846, "end": 170.82791346153846, "probability": 0.998046875 }, { "word": "下", "start": 170.82791346153846, "end": 170.98791346153845, "probability": 0.99365234375 }, { "word": "来", "start": 170.98791346153845, "end": 171.20791346153845, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.084868532940022, "compression_ratio": 1.3660714285714286, "no_speech_prob": 2.4770080084279833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但要是每生成一个token都重新计算所有这些表征浪费会大的离谱所以模型会把它们存下来这些存储下来的状态就是kb缓存token越多kb缓存就越大层数越多kb缓存也越大所以当上下文规模达到100万token时瓶颈就不再只是算力速度了kb缓存带来的延迟和成本问题会逐渐凸显你需要一次性计算搬运的token数量是注意力机制最初设计承载量的10到100倍所有欲填充" } ], "id": 83, "review_annotation": "", "row_id": 83, "bounded_row_id": 83, "parent_row_ids": [ 83 ], "parent_row_id": 83, "packet_id": 6, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 164.72791346153846, "packet_end": 193.91589903846156, "raw_text": "所以模型会把它们存下来", "selected_text": "所以模型会把它们存下来", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 83 ], "packet_id": 6, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 171.20791346153845, "end": 173.84791346153847, "text": "这些存储下来的状态就是kb缓存", "raw_asr_text": "这些存储下来的状态就是kb缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 6, "words": [ { "word": "这", "start": 171.20791346153845, "end": 171.66791346153846, "probability": 0.984375 }, { "word": "些", "start": 171.66791346153846, "end": 171.82791346153846, "probability": 0.99951171875 }, { "word": "存", "start": 171.82791346153846, "end": 172.02791346153847, "probability": 0.99755859375 }, { "word": "储", "start": 172.02791346153847, "end": 172.16791346153846, "probability": 0.99853515625 }, { "word": "下", "start": 172.16791346153846, "end": 172.32791346153846, "probability": 0.99658203125 }, { "word": "来", "start": 172.32791346153846, "end": 172.44791346153846, "probability": 0.99951171875 }, { "word": "的", "start": 172.44791346153846, "end": 172.50791346153846, "probability": 0.998046875 }, { "word": "状", "start": 172.50791346153846, "end": 172.66791346153846, "probability": 0.9970703125 }, { "word": "态", "start": 172.66791346153846, "end": 172.86791346153848, "probability": 1.0 }, { "word": "就是", "start": 172.86791346153848, "end": 173.18791346153847, "probability": 0.98486328125 }, { "word": "k", "start": 173.18791346153847, "end": 173.34791346153847, "probability": 0.54052734375 }, { "word": "b", "start": 173.34791346153847, "end": 173.44791346153846, "probability": 0.41015625 }, { "word": "缓", "start": 173.44791346153846, "end": 173.60791346153846, "probability": 0.99609375 }, { "word": "存", "start": 173.60791346153846, "end": 173.84791346153847, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.084868532940022, "compression_ratio": 1.3660714285714286, "no_speech_prob": 2.4770080084279833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但要是每生成一个token都重新计算所有这些表征浪费会大的离谱所以模型会把它们存下来这些存储下来的状态就是kb缓存token越多kb缓存就越大层数越多kb缓存也越大所以当上下文规模达到100万token时瓶颈就不再只是算力速度了kb缓存带来的延迟和成本问题会逐渐凸显你需要一次性计算搬运的token数量是注意力机制最初设计承载量的10到100倍所有欲填充" } ], "id": 84, "review_annotation": "", "row_id": 84, "bounded_row_id": 84, "parent_row_ids": [ 84 ], "parent_row_id": 84, "packet_id": 6, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 164.72791346153846, "packet_end": 193.91589903846156, "raw_text": "这些存储下来的状态就是kb缓存", "selected_text": "这些存储下来的状态就是kb缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 84 ], "packet_id": 6, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 173.84791346153847, "end": 176.30791346153848, "text": "token越多kb缓存就越大", "raw_asr_text": "token越多kb缓存就越大", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 6, "words": [ { "word": "t", "start": 173.84791346153847, "end": 174.24791346153847, "probability": 0.8857421875 }, { "word": "oken", "start": 174.24791346153847, "end": 174.44791346153846, "probability": 0.78955078125 }, { "word": "越", "start": 174.44791346153846, "end": 174.64791346153845, "probability": 0.75927734375 }, { "word": "多", "start": 174.64791346153845, "end": 174.92791346153845, "probability": 0.99951171875 }, { "word": "k", "start": 174.92791346153845, "end": 175.20791346153845, "probability": 0.81787109375 }, { "word": "b", "start": 175.20791346153845, "end": 175.36791346153848, "probability": 0.99853515625 }, { "word": "缓", "start": 175.36791346153848, "end": 175.54791346153846, "probability": 1.0 }, { "word": "存", "start": 175.54791346153846, "end": 175.74791346153847, "probability": 1.0 }, { "word": "就", "start": 175.74791346153847, "end": 175.90791346153847, "probability": 0.99462890625 }, { "word": "越", "start": 175.90791346153847, "end": 176.06791346153847, "probability": 0.994140625 }, { "word": "大", "start": 176.06791346153847, "end": 176.30791346153848, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.084868532940022, "compression_ratio": 1.3660714285714286, "no_speech_prob": 2.4770080084279833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但要是每生成一个token都重新计算所有这些表征浪费会大的离谱所以模型会把它们存下来这些存储下来的状态就是kb缓存token越多kb缓存就越大层数越多kb缓存也越大所以当上下文规模达到100万token时瓶颈就不再只是算力速度了kb缓存带来的延迟和成本问题会逐渐凸显你需要一次性计算搬运的token数量是注意力机制最初设计承载量的10到100倍所有欲填充" } ], "id": 85, "review_annotation": "", "row_id": 85, "bounded_row_id": 85, "parent_row_ids": [ 85 ], "parent_row_id": 85, "packet_id": 6, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 164.72791346153846, "packet_end": 193.91589903846156, "raw_text": "token越多kb缓存就越大", "selected_text": "token越多kb缓存就越大", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 85 ], "packet_id": 6, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 176.30791346153848, "end": 178.74791346153847, "text": "层数越多kb缓存也越大", "raw_asr_text": "层数越多kb缓存也越大", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 6, "words": [ { "word": "层", "start": 176.30791346153848, "end": 176.76791346153846, "probability": 0.96484375 }, { "word": "数", "start": 176.76791346153846, "end": 176.96791346153847, "probability": 0.998046875 }, { "word": "越", "start": 176.96791346153847, "end": 177.14791346153845, "probability": 0.99755859375 }, { "word": "多", "start": 177.14791346153845, "end": 177.38791346153846, "probability": 1.0 }, { "word": "k", "start": 177.38791346153846, "end": 177.76791346153846, "probability": 0.9931640625 }, { "word": "b", "start": 177.76791346153846, "end": 177.88791346153846, "probability": 0.99951171875 }, { "word": "缓", "start": 177.88791346153846, "end": 178.06791346153847, "probability": 1.0 }, { "word": "存", "start": 178.06791346153847, "end": 178.24791346153847, "probability": 0.99951171875 }, { "word": "也", "start": 178.24791346153847, "end": 178.38791346153846, "probability": 0.99462890625 }, { "word": "越", "start": 178.38791346153846, "end": 178.52791346153847, "probability": 0.98583984375 }, { "word": "大", "start": 178.52791346153847, "end": 178.74791346153847, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.084868532940022, "compression_ratio": 1.3660714285714286, "no_speech_prob": 2.4770080084279833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但要是每生成一个token都重新计算所有这些表征浪费会大的离谱所以模型会把它们存下来这些存储下来的状态就是kb缓存token越多kb缓存就越大层数越多kb缓存也越大所以当上下文规模达到100万token时瓶颈就不再只是算力速度了kb缓存带来的延迟和成本问题会逐渐凸显你需要一次性计算搬运的token数量是注意力机制最初设计承载量的10到100倍所有欲填充" } ], "id": 86, "review_annotation": "", "row_id": 86, "bounded_row_id": 86, "parent_row_ids": [ 86 ], "parent_row_id": 86, "packet_id": 6, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 164.72791346153846, "packet_end": 193.91589903846156, "raw_text": "层数越多kb缓存也越大", "selected_text": "层数越多kb缓存也越大", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 86 ], "packet_id": 6, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 178.74791346153847, "end": 181.50791346153846, "text": "所以当上下文规模达到100万token时", "raw_asr_text": "所以当上下文规模达到100万token时", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 6, "words": [ { "word": "所以", "start": 178.74791346153847, "end": 179.24791346153847, "probability": 0.98779296875 }, { "word": "当", "start": 179.24791346153847, "end": 179.40791346153847, "probability": 0.9892578125 }, { "word": "上", "start": 179.40791346153847, "end": 179.58791346153845, "probability": 0.998046875 }, { "word": "下", "start": 179.58791346153845, "end": 179.72791346153846, "probability": 0.9970703125 }, { "word": "文", "start": 179.72791346153846, "end": 179.86791346153848, "probability": 0.91357421875 }, { "word": "规", "start": 179.86791346153848, "end": 179.98791346153845, "probability": 0.9990234375 }, { "word": "模", "start": 179.98791346153845, "end": 180.18791346153847, "probability": 0.99951171875 }, { "word": "达", "start": 180.18791346153847, "end": 180.36791346153848, "probability": 0.9931640625 }, { "word": "到", "start": 180.36791346153848, "end": 180.48791346153845, "probability": 0.9931640625 }, { "word": "100", "start": 180.48791346153845, "end": 180.70791346153845, "probability": 0.939453125 }, { "word": "万", "start": 180.70791346153845, "end": 180.90791346153847, "probability": 0.9755859375 }, { "word": "t", "start": 180.90791346153847, "end": 181.04791346153846, "probability": 0.978515625 }, { "word": "oken", "start": 181.04791346153846, "end": 181.22791346153846, "probability": 0.99658203125 }, { "word": "时", "start": 181.22791346153846, "end": 181.50791346153846, "probability": 0.9814453125 } ], "asr_metrics": { "avg_logprob": -0.084868532940022, "compression_ratio": 1.3660714285714286, "no_speech_prob": 2.4770080084279833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但要是每生成一个token都重新计算所有这些表征浪费会大的离谱所以模型会把它们存下来这些存储下来的状态就是kb缓存token越多kb缓存就越大层数越多kb缓存也越大所以当上下文规模达到100万token时瓶颈就不再只是算力速度了kb缓存带来的延迟和成本问题会逐渐凸显你需要一次性计算搬运的token数量是注意力机制最初设计承载量的10到100倍所有欲填充" } ], "id": 87, "review_annotation": "", "row_id": 87, "bounded_row_id": 87, "parent_row_ids": [ 87 ], "parent_row_id": 87, "packet_id": 6, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 164.72791346153846, "packet_end": 193.91589903846156, "raw_text": "所以当上下文规模达到100万token时", "selected_text": "所以当上下文规模达到100万token时", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 87 ], "packet_id": 6, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 181.50791346153846, "end": 183.94791346153846, "text": "瓶颈就不再只是算力速度了", "raw_asr_text": "瓶颈就不再只是算力速度了", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 6, "words": [ { "word": "瓶", "start": 181.50791346153846, "end": 181.96791346153847, "probability": 0.9951171875 }, { "word": "颈", "start": 181.96791346153847, "end": 182.18791346153847, "probability": 0.986328125 }, { "word": "就", "start": 182.18791346153847, "end": 182.38791346153846, "probability": 0.98974609375 }, { "word": "不", "start": 182.38791346153846, "end": 182.64791346153845, "probability": 0.99755859375 }, { "word": "再", "start": 182.64791346153845, "end": 182.82791346153846, "probability": 0.99267578125 }, { "word": "只是", "start": 182.82791346153846, "end": 183.12791346153847, "probability": 0.9775390625 }, { "word": "算", "start": 183.12791346153847, "end": 183.34791346153847, "probability": 0.9931640625 }, { "word": "力", "start": 183.34791346153847, "end": 183.46791346153847, "probability": 0.9853515625 }, { "word": "速", "start": 183.46791346153847, "end": 183.64791346153845, "probability": 0.98486328125 }, { "word": "度", "start": 183.64791346153845, "end": 183.80791346153848, "probability": 1.0 }, { "word": "了", "start": 183.80791346153848, "end": 183.94791346153846, "probability": 0.9970703125 } ], "asr_metrics": { "avg_logprob": -0.084868532940022, "compression_ratio": 1.3660714285714286, "no_speech_prob": 2.4770080084279833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但要是每生成一个token都重新计算所有这些表征浪费会大的离谱所以模型会把它们存下来这些存储下来的状态就是kb缓存token越多kb缓存就越大层数越多kb缓存也越大所以当上下文规模达到100万token时瓶颈就不再只是算力速度了kb缓存带来的延迟和成本问题会逐渐凸显你需要一次性计算搬运的token数量是注意力机制最初设计承载量的10到100倍所有欲填充" } ], "id": 88, "review_annotation": "", "row_id": 88, "bounded_row_id": 88, "parent_row_ids": [ 88 ], "parent_row_id": 88, "packet_id": 6, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 164.72791346153846, "packet_end": 193.91589903846156, "raw_text": "瓶颈就不再只是算力速度了", "selected_text": "瓶颈就不再只是算力速度了", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 88 ], "packet_id": 6, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 183.94791346153846, "end": 187.02791346153847, "text": "kb缓存带来的延迟和成本问题会逐渐凸显", "raw_asr_text": "kb缓存带来的延迟和成本问题会逐渐凸显", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 6, "words": [ { "word": "k", "start": 183.94791346153846, "end": 184.26791346153846, "probability": 0.98388671875 }, { "word": "b", "start": 184.26791346153846, "end": 184.40791346153847, "probability": 0.99853515625 }, { "word": "缓", "start": 184.40791346153847, "end": 184.56791346153847, "probability": 1.0 }, { "word": "存", "start": 184.56791346153847, "end": 184.70791346153845, "probability": 0.99951171875 }, { "word": "带", "start": 184.70791346153845, "end": 184.86791346153848, "probability": 0.99072265625 }, { "word": "来", "start": 184.86791346153848, "end": 184.98791346153845, "probability": 0.99951171875 }, { "word": "的", "start": 184.98791346153845, "end": 185.06791346153847, "probability": 0.99609375 }, { "word": "延", "start": 185.06791346153847, "end": 185.28791346153847, "probability": 0.9990234375 }, { "word": "迟", "start": 185.28791346153847, "end": 185.46791346153847, "probability": 0.9970703125 }, { "word": "和", "start": 185.46791346153847, "end": 185.58791346153845, "probability": 0.994140625 }, { "word": "成", "start": 185.58791346153845, "end": 185.74791346153847, "probability": 0.9990234375 }, { "word": "本", "start": 185.74791346153847, "end": 185.86791346153848, "probability": 0.99951171875 }, { "word": "问题", "start": 185.86791346153848, "end": 186.12791346153847, "probability": 0.99853515625 }, { "word": "会", "start": 186.12791346153847, "end": 186.34791346153847, "probability": 0.984375 }, { "word": "逐", "start": 186.34791346153847, "end": 186.50791346153846, "probability": 0.99951171875 }, { "word": "渐", "start": 186.50791346153846, "end": 186.66791346153846, "probability": 0.9990234375 }, { "word": "凸", "start": 186.66791346153846, "end": 186.80791346153848, "probability": 0.9599609375 }, { "word": "显", "start": 186.80791346153848, "end": 187.02791346153847, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.084868532940022, "compression_ratio": 1.3660714285714286, "no_speech_prob": 2.4770080084279833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但要是每生成一个token都重新计算所有这些表征浪费会大的离谱所以模型会把它们存下来这些存储下来的状态就是kb缓存token越多kb缓存就越大层数越多kb缓存也越大所以当上下文规模达到100万token时瓶颈就不再只是算力速度了kb缓存带来的延迟和成本问题会逐渐凸显你需要一次性计算搬运的token数量是注意力机制最初设计承载量的10到100倍所有欲填充" } ], "id": 89, "review_annotation": "", "row_id": 89, "bounded_row_id": 89, "parent_row_ids": [ 89 ], "parent_row_id": 89, "packet_id": 6, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 164.72791346153846, "packet_end": 193.91589903846156, "raw_text": "kb缓存带来的延迟和成本问题会逐渐凸显", "selected_text": "kb缓存带来的延迟和成本问题会逐渐凸显", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 89 ], "packet_id": 6, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 187.02791346153847, "end": 189.46791346153847, "text": "你需要一次性计算搬运的token数量", "raw_asr_text": "你需要一次性计算搬运的token数量", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 6, "words": [ { "word": "你", "start": 187.02791346153847, "end": 187.28791346153847, "probability": 0.98046875 }, { "word": "需要", "start": 187.28791346153847, "end": 187.44791346153846, "probability": 0.99755859375 }, { "word": "一次", "start": 187.44791346153846, "end": 187.70791346153845, "probability": 0.998046875 }, { "word": "性", "start": 187.70791346153845, "end": 187.92791346153845, "probability": 0.9990234375 }, { "word": "计", "start": 187.92791346153845, "end": 188.06791346153847, "probability": 1.0 }, { "word": "算", "start": 188.06791346153847, "end": 188.22791346153846, "probability": 1.0 }, { "word": "搬", "start": 188.22791346153846, "end": 188.40791346153847, "probability": 0.99462890625 }, { "word": "运", "start": 188.40791346153847, "end": 188.56791346153847, "probability": 0.998046875 }, { "word": "的", "start": 188.56791346153847, "end": 188.68791346153847, "probability": 0.99755859375 }, { "word": "t", "start": 188.68791346153847, "end": 188.82791346153846, "probability": 0.98828125 }, { "word": "oken", "start": 188.82791346153846, "end": 188.98791346153845, "probability": 0.9609375 }, { "word": "数", "start": 188.98791346153845, "end": 189.24791346153847, "probability": 0.9609375 }, { "word": "量", "start": 189.24791346153847, "end": 189.46791346153847, "probability": 0.99609375 } ], "asr_metrics": { "avg_logprob": -0.084868532940022, "compression_ratio": 1.3660714285714286, "no_speech_prob": 2.4770080084279833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但要是每生成一个token都重新计算所有这些表征浪费会大的离谱所以模型会把它们存下来这些存储下来的状态就是kb缓存token越多kb缓存就越大层数越多kb缓存也越大所以当上下文规模达到100万token时瓶颈就不再只是算力速度了kb缓存带来的延迟和成本问题会逐渐凸显你需要一次性计算搬运的token数量是注意力机制最初设计承载量的10到100倍所有欲填充" } ], "id": 90, "review_annotation": "", "row_id": 90, "bounded_row_id": 90, "parent_row_ids": [ 90 ], "parent_row_id": 90, "packet_id": 6, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 164.72791346153846, "packet_end": 193.91589903846156, "raw_text": "你需要一次性计算搬运的token数量", "selected_text": "你需要一次性计算搬运的token数量", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 90 ], "packet_id": 6, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 189.46791346153847, "end": 192.72791346153846, "text": "是注意力机制最初设计承载量的10到100倍", "raw_asr_text": "是注意力机制最初设计承载量的10到100倍", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 6, "words": [ { "word": "是", "start": 189.46791346153847, "end": 189.74791346153847, "probability": 0.98486328125 }, { "word": "注意", "start": 189.74791346153847, "end": 190.00791346153846, "probability": 0.978515625 }, { "word": "力", "start": 190.00791346153846, "end": 190.16791346153846, "probability": 0.99609375 }, { "word": "机", "start": 190.16791346153846, "end": 190.30791346153848, "probability": 0.98291015625 }, { "word": "制", "start": 190.30791346153848, "end": 190.46791346153847, "probability": 0.9892578125 }, { "word": "最", "start": 190.46791346153847, "end": 190.74791346153847, "probability": 0.98486328125 }, { "word": "初", "start": 190.74791346153847, "end": 190.96791346153847, "probability": 0.99658203125 }, { "word": "设", "start": 190.96791346153847, "end": 191.20791346153845, "probability": 0.9970703125 }, { "word": "计", "start": 191.20791346153845, "end": 191.32791346153846, "probability": 1.0 }, { "word": "承", "start": 191.32791346153846, "end": 191.52791346153847, "probability": 0.994140625 }, { "word": "载", "start": 191.52791346153847, "end": 191.68791346153847, "probability": 0.9990234375 }, { "word": "量", "start": 191.68791346153847, "end": 191.84791346153847, "probability": 0.99609375 }, { "word": "的", "start": 191.84791346153847, "end": 191.94791346153846, "probability": 0.98681640625 }, { "word": "10", "start": 191.94791346153846, "end": 192.14791346153845, "probability": 0.935546875 }, { "word": "到", "start": 192.14791346153845, "end": 192.24791346153847, "probability": 0.51123046875 }, { "word": "100", "start": 192.24791346153847, "end": 192.48791346153845, "probability": 0.99658203125 }, { "word": "倍", "start": 192.48791346153845, "end": 192.72791346153846, "probability": 0.99560546875 } ], "asr_metrics": { "avg_logprob": -0.084868532940022, "compression_ratio": 1.3660714285714286, "no_speech_prob": 2.4770080084279833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但要是每生成一个token都重新计算所有这些表征浪费会大的离谱所以模型会把它们存下来这些存储下来的状态就是kb缓存token越多kb缓存就越大层数越多kb缓存也越大所以当上下文规模达到100万token时瓶颈就不再只是算力速度了kb缓存带来的延迟和成本问题会逐渐凸显你需要一次性计算搬运的token数量是注意力机制最初设计承载量的10到100倍所有欲填充" } ], "id": 91, "review_annotation": "", "row_id": 91, "bounded_row_id": 91, "parent_row_ids": [ 91 ], "parent_row_id": 91, "packet_id": 6, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 164.72791346153846, "packet_end": 193.91589903846156, "raw_text": "是注意力机制最初设计承载量的10到100倍", "selected_text": "是注意力机制最初设计承载量的10到100倍", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 91 ], "packet_id": 6, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 193.11589903846155, "end": 194.27589903846155, "text": "所有预天穷算力", "raw_asr_text": "所有预天穷算力", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 7, "words": [ { "word": "所有", "start": 192.89589903846155, "end": 193.23589903846153, "probability": 0.454345703125 }, { "word": "预", "start": 193.23589903846153, "end": 193.51589903846153, "probability": 0.6533203125 }, { "word": "天", "start": 193.51589903846153, "end": 193.67589903846155, "probability": 0.90185546875 }, { "word": "穷", "start": 193.67589903846155, "end": 193.85589903846153, "probability": 0.544921875 }, { "word": "算", "start": 193.85589903846153, "end": 194.03589903846154, "probability": 0.9716796875 }, { "word": "力", "start": 194.03589903846154, "end": 194.27589903846155, "probability": 0.990234375 } ], "asr_metrics": { "avg_logprob": -0.11731879699406365, "compression_ratio": 1.234375, "no_speech_prob": 5.1019511532190975e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有预天穷算力HBM SSD 还有带宽如今都得适配这类需求这是因为我们大量使用智能体工具至少也是当下的模型服务模式导致的V4.1的优化思路从这里开始就很有意思了因为在V4版本中他们已经用压缩吸收注意力CSA将多个Token压缩为更少的KV条目之后再用吸收索引器判断这些压缩后的内存里哪些部分真正值得读取因此为了应对不断增长的上下文长度他们已经采用了" } ], "id": 92, "review_annotation": "", "row_id": 92, "bounded_row_id": 92, "parent_row_ids": [ 92 ], "parent_row_id": 92, "packet_id": 7, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 192.31589903846154, "packet_end": 221.50388461538464, "raw_text": "所有预天穷算力", "selected_text": "所有预天穷算力", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 92 ], "packet_id": 7, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 194.27589903846155, "end": 196.79589903846153, "text": "HBM SSD 还有带宽", "raw_asr_text": "HBM SSD 还有带宽", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 7, "words": [ { "word": "H", "start": 194.27589903846155, "end": 194.73589903846153, "probability": 0.59912109375 }, { "word": "BM", "start": 194.73589903846153, "end": 195.09589903846154, "probability": 0.94189453125 }, { "word": " SSD", "start": 195.09589903846154, "end": 195.81589903846154, "probability": 0.5810546875 }, { "word": " 还有", "start": 196.09589903846154, "end": 196.27589903846155, "probability": 0.91943359375 }, { "word": "带", "start": 196.27589903846155, "end": 196.57589903846153, "probability": 0.958984375 }, { "word": "宽", "start": 196.57589903846153, "end": 196.79589903846153, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.11731879699406365, "compression_ratio": 1.234375, "no_speech_prob": 5.1019511532190975e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有预天穷算力HBM SSD 还有带宽如今都得适配这类需求这是因为我们大量使用智能体工具至少也是当下的模型服务模式导致的V4.1的优化思路从这里开始就很有意思了因为在V4版本中他们已经用压缩吸收注意力CSA将多个Token压缩为更少的KV条目之后再用吸收索引器判断这些压缩后的内存里哪些部分真正值得读取因此为了应对不断增长的上下文长度他们已经采用了" } ], "id": 93, "review_annotation": "", "row_id": 93, "bounded_row_id": 93, "parent_row_ids": [ 93 ], "parent_row_id": 93, "packet_id": 7, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 192.31589903846154, "packet_end": 221.50388461538464, "raw_text": "HBM SSD 还有带宽", "selected_text": "HBM SSD 还有带宽", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 93 ], "packet_id": 7, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 196.79589903846153, "end": 198.83589903846155, "text": "如今都得适配这类需求", "raw_asr_text": "如今都得适配这类需求", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 7, "words": [ { "word": "如", "start": 196.79589903846153, "end": 197.11589903846155, "probability": 0.96728515625 }, { "word": "今", "start": 197.11589903846155, "end": 197.39589903846155, "probability": 0.97900390625 }, { "word": "都", "start": 197.39589903846155, "end": 197.61589903846155, "probability": 0.98193359375 }, { "word": "得", "start": 197.61589903846155, "end": 197.77589903846155, "probability": 0.98583984375 }, { "word": "适", "start": 197.77589903846155, "end": 197.99589903846154, "probability": 0.9951171875 }, { "word": "配", "start": 197.99589903846154, "end": 198.15589903846154, "probability": 0.99853515625 }, { "word": "这", "start": 198.15589903846154, "end": 198.35589903846153, "probability": 0.9892578125 }, { "word": "类", "start": 198.35589903846153, "end": 198.49589903846154, "probability": 0.998046875 }, { "word": "需", "start": 198.49589903846154, "end": 198.63589903846153, "probability": 0.9951171875 }, { "word": "求", "start": 198.63589903846153, "end": 198.83589903846155, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.11731879699406365, "compression_ratio": 1.234375, "no_speech_prob": 5.1019511532190975e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有预天穷算力HBM SSD 还有带宽如今都得适配这类需求这是因为我们大量使用智能体工具至少也是当下的模型服务模式导致的V4.1的优化思路从这里开始就很有意思了因为在V4版本中他们已经用压缩吸收注意力CSA将多个Token压缩为更少的KV条目之后再用吸收索引器判断这些压缩后的内存里哪些部分真正值得读取因此为了应对不断增长的上下文长度他们已经采用了" } ], "id": 94, "review_annotation": "", "row_id": 94, "bounded_row_id": 94, "parent_row_ids": [ 94 ], "parent_row_id": 94, "packet_id": 7, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 192.31589903846154, "packet_end": 221.50388461538464, "raw_text": "如今都得适配这类需求", "selected_text": "如今都得适配这类需求", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 94 ], "packet_id": 7, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 198.83589903846155, "end": 201.13589903846153, "text": "这是因为我们大量使用智能体工具", "raw_asr_text": "这是因为我们大量使用智能体工具", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 7, "words": [ { "word": "这是", "start": 198.83589903846155, "end": 199.17589903846155, "probability": 0.92919921875 }, { "word": "因为", "start": 199.17589903846155, "end": 199.43589903846154, "probability": 0.990234375 }, { "word": "我们", "start": 199.43589903846154, "end": 199.65589903846154, "probability": 0.98828125 }, { "word": "大", "start": 199.65589903846154, "end": 199.83589903846155, "probability": 0.994140625 }, { "word": "量", "start": 199.83589903846155, "end": 199.99589903846154, "probability": 0.9990234375 }, { "word": "使", "start": 199.99589903846154, "end": 200.21589903846154, "probability": 0.990234375 }, { "word": "用", "start": 200.21589903846154, "end": 200.31589903846154, "probability": 1.0 }, { "word": "智", "start": 200.31589903846154, "end": 200.51589903846153, "probability": 0.998046875 }, { "word": "能", "start": 200.51589903846153, "end": 200.61589903846155, "probability": 0.99951171875 }, { "word": "体", "start": 200.61589903846155, "end": 200.77589903846155, "probability": 0.9951171875 }, { "word": "工", "start": 200.77589903846155, "end": 200.99589903846154, "probability": 0.982421875 }, { "word": "具", "start": 200.99589903846154, "end": 201.13589903846153, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11731879699406365, "compression_ratio": 1.234375, "no_speech_prob": 5.1019511532190975e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有预天穷算力HBM SSD 还有带宽如今都得适配这类需求这是因为我们大量使用智能体工具至少也是当下的模型服务模式导致的V4.1的优化思路从这里开始就很有意思了因为在V4版本中他们已经用压缩吸收注意力CSA将多个Token压缩为更少的KV条目之后再用吸收索引器判断这些压缩后的内存里哪些部分真正值得读取因此为了应对不断增长的上下文长度他们已经采用了" } ], "id": 95, "review_annotation": "", "row_id": 95, "bounded_row_id": 95, "parent_row_ids": [ 95 ], "parent_row_id": 95, "packet_id": 7, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 192.31589903846154, "packet_end": 221.50388461538464, "raw_text": "这是因为我们大量使用智能体工具", "selected_text": "这是因为我们大量使用智能体工具", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 95 ], "packet_id": 7, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 201.13589903846153, "end": 203.89589903846155, "text": "至少也是当下的模型服务模式导致的", "raw_asr_text": "至少也是当下的模型服务模式导致的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 7, "words": [ { "word": "至", "start": 201.13589903846153, "end": 201.39589903846155, "probability": 0.98486328125 }, { "word": "少", "start": 201.39589903846155, "end": 201.63589903846153, "probability": 0.9990234375 }, { "word": "也是", "start": 201.63589903846153, "end": 201.89589903846155, "probability": 0.9853515625 }, { "word": "当", "start": 201.89589903846155, "end": 202.13589903846153, "probability": 0.9931640625 }, { "word": "下", "start": 202.13589903846153, "end": 202.29589903846153, "probability": 0.99951171875 }, { "word": "的", "start": 202.29589903846153, "end": 202.41589903846153, "probability": 0.986328125 }, { "word": "模", "start": 202.41589903846153, "end": 202.57589903846153, "probability": 0.9951171875 }, { "word": "型", "start": 202.57589903846153, "end": 202.73589903846153, "probability": 0.99951171875 }, { "word": "服", "start": 202.73589903846153, "end": 202.93589903846154, "probability": 0.9970703125 }, { "word": "务", "start": 202.93589903846154, "end": 203.07589903846153, "probability": 1.0 }, { "word": "模", "start": 203.07589903846153, "end": 203.23589903846153, "probability": 0.99560546875 }, { "word": "式", "start": 203.23589903846153, "end": 203.43589903846154, "probability": 1.0 }, { "word": "导", "start": 203.43589903846154, "end": 203.61589903846155, "probability": 1.0 }, { "word": "致", "start": 203.61589903846155, "end": 203.77589903846155, "probability": 1.0 }, { "word": "的", "start": 203.77589903846155, "end": 203.89589903846155, "probability": 0.98388671875 } ], "asr_metrics": { "avg_logprob": -0.11731879699406365, "compression_ratio": 1.234375, "no_speech_prob": 5.1019511532190975e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有预天穷算力HBM SSD 还有带宽如今都得适配这类需求这是因为我们大量使用智能体工具至少也是当下的模型服务模式导致的V4.1的优化思路从这里开始就很有意思了因为在V4版本中他们已经用压缩吸收注意力CSA将多个Token压缩为更少的KV条目之后再用吸收索引器判断这些压缩后的内存里哪些部分真正值得读取因此为了应对不断增长的上下文长度他们已经采用了" } ], "id": 96, "review_annotation": "", "row_id": 96, "bounded_row_id": 96, "parent_row_ids": [ 96 ], "parent_row_id": 96, "packet_id": 7, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 192.31589903846154, "packet_end": 221.50388461538464, "raw_text": "至少也是当下的模型服务模式导致的", "selected_text": "至少也是当下的模型服务模式导致的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 96 ], "packet_id": 7, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 203.89589903846155, "end": 206.65589903846154, "text": "V4.1的优化思路从这里开始就很有意思了", "raw_asr_text": "V4.1的优化思路从这里开始就很有意思了", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 7, "words": [ { "word": "V", "start": 203.89589903846155, "end": 204.23589903846153, "probability": 0.81591796875 }, { "word": "4.", "start": 204.23589903846153, "end": 204.37589903846154, "probability": 0.89892578125 }, { "word": "1", "start": 204.47589903846153, "end": 204.63589903846153, "probability": 0.9609375 }, { "word": "的", "start": 204.63589903846153, "end": 204.75589903846154, "probability": 0.90087890625 }, { "word": "优", "start": 204.75589903846154, "end": 204.83589903846155, "probability": 0.8828125 }, { "word": "化", "start": 204.83589903846155, "end": 204.97589903846153, "probability": 0.99365234375 }, { "word": "思", "start": 204.97589903846153, "end": 205.09589903846154, "probability": 0.99755859375 }, { "word": "路", "start": 205.09589903846154, "end": 205.19589903846153, "probability": 0.99462890625 }, { "word": "从", "start": 205.19589903846153, "end": 205.35589903846153, "probability": 0.9853515625 }, { "word": "这里", "start": 205.35589903846153, "end": 205.59589903846154, "probability": 0.9794921875 }, { "word": "开始", "start": 205.59589903846154, "end": 205.87589903846154, "probability": 0.994140625 }, { "word": "就", "start": 205.87589903846154, "end": 206.03589903846154, "probability": 0.98876953125 }, { "word": "很有", "start": 206.03589903846154, "end": 206.25589903846154, "probability": 0.9873046875 }, { "word": "意思", "start": 206.25589903846154, "end": 206.51589903846153, "probability": 1.0 }, { "word": "了", "start": 206.51589903846153, "end": 206.65589903846154, "probability": 0.95556640625 } ], "asr_metrics": { "avg_logprob": -0.11731879699406365, "compression_ratio": 1.234375, "no_speech_prob": 5.1019511532190975e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有预天穷算力HBM SSD 还有带宽如今都得适配这类需求这是因为我们大量使用智能体工具至少也是当下的模型服务模式导致的V4.1的优化思路从这里开始就很有意思了因为在V4版本中他们已经用压缩吸收注意力CSA将多个Token压缩为更少的KV条目之后再用吸收索引器判断这些压缩后的内存里哪些部分真正值得读取因此为了应对不断增长的上下文长度他们已经采用了" } ], "id": 97, "review_annotation": "", "row_id": 97, "bounded_row_id": 97, "parent_row_ids": [ 97 ], "parent_row_id": 97, "packet_id": 7, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 192.31589903846154, "packet_end": 221.50388461538464, "raw_text": "V4.1的优化思路从这里开始就很有意思了", "selected_text": "V4.1的优化思路从这里开始就很有意思了", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 97 ], "packet_id": 7, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 206.65589903846154, "end": 207.65589903846154, "text": "因为在V4版本中", "raw_asr_text": "因为在V4版本中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 7, "words": [ { "word": "因为", "start": 206.65589903846154, "end": 206.83589903846155, "probability": 0.966796875 }, { "word": "在", "start": 206.83589903846155, "end": 206.97589903846153, "probability": 0.990234375 }, { "word": "V", "start": 206.97589903846153, "end": 207.09589903846154, "probability": 0.9765625 }, { "word": "4", "start": 207.09589903846154, "end": 207.19589903846153, "probability": 0.98046875 }, { "word": "版", "start": 207.19589903846153, "end": 207.37589903846154, "probability": 0.95166015625 }, { "word": "本", "start": 207.37589903846154, "end": 207.55589903846155, "probability": 0.9990234375 }, { "word": "中", "start": 207.55589903846155, "end": 207.65589903846154, "probability": 0.98828125 } ], "asr_metrics": { "avg_logprob": -0.11731879699406365, "compression_ratio": 1.234375, "no_speech_prob": 5.1019511532190975e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有预天穷算力HBM SSD 还有带宽如今都得适配这类需求这是因为我们大量使用智能体工具至少也是当下的模型服务模式导致的V4.1的优化思路从这里开始就很有意思了因为在V4版本中他们已经用压缩吸收注意力CSA将多个Token压缩为更少的KV条目之后再用吸收索引器判断这些压缩后的内存里哪些部分真正值得读取因此为了应对不断增长的上下文长度他们已经采用了" } ], "id": 98, "review_annotation": "", "row_id": 98, "bounded_row_id": 98, "parent_row_ids": [ 98 ], "parent_row_id": 98, "packet_id": 7, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 192.31589903846154, "packet_end": 221.50388461538464, "raw_text": "因为在V4版本中", "selected_text": "因为在V4版本中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 98 ], "packet_id": 7, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 207.65589903846154, "end": 209.35589903846153, "text": "他们已经用压缩吸收注意力", "raw_asr_text": "他们已经用压缩吸收注意力", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 7, "words": [ { "word": "他们", "start": 207.65589903846154, "end": 207.91589903846153, "probability": 0.84619140625 }, { "word": "已经", "start": 207.91589903846153, "end": 208.13589903846153, "probability": 0.99560546875 }, { "word": "用", "start": 208.13589903846153, "end": 208.31589903846154, "probability": 0.994140625 }, { "word": "压", "start": 208.31589903846154, "end": 208.53589903846154, "probability": 0.9951171875 }, { "word": "缩", "start": 208.53589903846154, "end": 208.69589903846153, "probability": 0.9970703125 }, { "word": "吸", "start": 208.69589903846153, "end": 208.83589903846155, "probability": 0.481689453125 }, { "word": "收", "start": 208.83589903846155, "end": 208.97589903846153, "probability": 0.5224609375 }, { "word": "注意", "start": 208.97589903846153, "end": 209.21589903846154, "probability": 0.98974609375 }, { "word": "力", "start": 209.21589903846154, "end": 209.35589903846153, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.11731879699406365, "compression_ratio": 1.234375, "no_speech_prob": 5.1019511532190975e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有预天穷算力HBM SSD 还有带宽如今都得适配这类需求这是因为我们大量使用智能体工具至少也是当下的模型服务模式导致的V4.1的优化思路从这里开始就很有意思了因为在V4版本中他们已经用压缩吸收注意力CSA将多个Token压缩为更少的KV条目之后再用吸收索引器判断这些压缩后的内存里哪些部分真正值得读取因此为了应对不断增长的上下文长度他们已经采用了" } ], "id": 99, "review_annotation": "", "row_id": 99, "bounded_row_id": 99, "parent_row_ids": [ 99 ], "parent_row_id": 99, "packet_id": 7, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 192.31589903846154, "packet_end": 221.50388461538464, "raw_text": "他们已经用压缩吸收注意力", "selected_text": "他们已经用压缩吸收注意力", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 99 ], "packet_id": 7, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 209.35589903846153, "end": 212.47589903846153, "text": "CSA将多个Token压缩为更少的KV条目", "raw_asr_text": "CSA将多个Token压缩为更少的KV条目", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 7, "words": [ { "word": "C", "start": 209.35589903846153, "end": 209.55589903846155, "probability": 0.947265625 }, { "word": "SA", "start": 209.55589903846155, "end": 209.83589903846155, "probability": 0.98583984375 }, { "word": "将", "start": 209.83589903846155, "end": 210.01589903846153, "probability": 0.9794921875 }, { "word": "多", "start": 210.01589903846153, "end": 210.25589903846154, "probability": 0.99560546875 }, { "word": "个", "start": 210.25589903846154, "end": 210.43589903846154, "probability": 0.9873046875 }, { "word": "T", "start": 210.43589903846154, "end": 210.57589903846153, "probability": 0.4130859375 }, { "word": "oken", "start": 210.57589903846153, "end": 210.75589903846154, "probability": 0.599609375 }, { "word": "压", "start": 210.75589903846154, "end": 211.11589903846155, "probability": 0.8173828125 }, { "word": "缩", "start": 211.11589903846155, "end": 211.21589903846154, "probability": 1.0 }, { "word": "为", "start": 211.21589903846154, "end": 211.33589903846155, "probability": 0.99462890625 }, { "word": "更", "start": 211.33589903846155, "end": 211.51589903846153, "probability": 0.9990234375 }, { "word": "少", "start": 211.51589903846153, "end": 211.71589903846154, "probability": 0.9970703125 }, { "word": "的", "start": 211.71589903846154, "end": 211.83589903846155, "probability": 0.99267578125 }, { "word": "K", "start": 211.83589903846155, "end": 211.99589903846154, "probability": 0.90966796875 }, { "word": "V", "start": 211.99589903846154, "end": 212.11589903846155, "probability": 0.9931640625 }, { "word": "条", "start": 212.11589903846155, "end": 212.25589903846154, "probability": 0.9892578125 }, { "word": "目", "start": 212.25589903846154, "end": 212.47589903846153, "probability": 0.947265625 } ], "asr_metrics": { "avg_logprob": -0.11731879699406365, "compression_ratio": 1.234375, "no_speech_prob": 5.1019511532190975e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有预天穷算力HBM SSD 还有带宽如今都得适配这类需求这是因为我们大量使用智能体工具至少也是当下的模型服务模式导致的V4.1的优化思路从这里开始就很有意思了因为在V4版本中他们已经用压缩吸收注意力CSA将多个Token压缩为更少的KV条目之后再用吸收索引器判断这些压缩后的内存里哪些部分真正值得读取因此为了应对不断增长的上下文长度他们已经采用了" } ], "id": 100, "review_annotation": "", "row_id": 100, "bounded_row_id": 100, "parent_row_ids": [ 100 ], "parent_row_id": 100, "packet_id": 7, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 192.31589903846154, "packet_end": 221.50388461538464, "raw_text": "CSA将多个Token压缩为更少的KV条目", "selected_text": "CSA将多个Token压缩为更少的KV条目", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 100 ], "packet_id": 7, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 212.47589903846153, "end": 214.31589903846154, "text": "之后再用吸收索引器", "raw_asr_text": "之后再用吸收索引器", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 7, "words": [ { "word": "之", "start": 212.47589903846153, "end": 213.07589903846153, "probability": 0.994140625 }, { "word": "后", "start": 213.07589903846153, "end": 213.23589903846153, "probability": 1.0 }, { "word": "再", "start": 213.23589903846153, "end": 213.39589903846155, "probability": 0.96826171875 }, { "word": "用", "start": 213.39589903846155, "end": 213.53589903846154, "probability": 0.998046875 }, { "word": "吸", "start": 213.53589903846154, "end": 213.73589903846153, "probability": 0.9912109375 }, { "word": "收", "start": 213.73589903846153, "end": 213.89589903846155, "probability": 0.974609375 }, { "word": "索", "start": 213.89589903846155, "end": 214.01589903846153, "probability": 0.94775390625 }, { "word": "引", "start": 214.01589903846153, "end": 214.15589903846154, "probability": 0.9970703125 }, { "word": "器", "start": 214.15589903846154, "end": 214.31589903846154, "probability": 0.94970703125 } ], "asr_metrics": { "avg_logprob": -0.11731879699406365, "compression_ratio": 1.234375, "no_speech_prob": 5.1019511532190975e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有预天穷算力HBM SSD 还有带宽如今都得适配这类需求这是因为我们大量使用智能体工具至少也是当下的模型服务模式导致的V4.1的优化思路从这里开始就很有意思了因为在V4版本中他们已经用压缩吸收注意力CSA将多个Token压缩为更少的KV条目之后再用吸收索引器判断这些压缩后的内存里哪些部分真正值得读取因此为了应对不断增长的上下文长度他们已经采用了" } ], "id": 101, "review_annotation": "", "row_id": 101, "bounded_row_id": 101, "parent_row_ids": [ 101 ], "parent_row_id": 101, "packet_id": 7, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 192.31589903846154, "packet_end": 221.50388461538464, "raw_text": "之后再用吸收索引器", "selected_text": "之后再用吸收索引器", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 101 ], "packet_id": 7, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 214.31589903846154, "end": 216.01589903846153, "text": "判断这些压缩后的内存里", "raw_asr_text": "判断这些压缩后的内存里", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 7, "words": [ { "word": "判", "start": 214.31589903846154, "end": 214.59589903846154, "probability": 0.99755859375 }, { "word": "断", "start": 214.59589903846154, "end": 214.77589903846155, "probability": 1.0 }, { "word": "这", "start": 214.77589903846155, "end": 214.89589903846155, "probability": 0.994140625 }, { "word": "些", "start": 214.89589903846155, "end": 215.03589903846154, "probability": 0.99951171875 }, { "word": "压", "start": 215.03589903846154, "end": 215.19589903846153, "probability": 1.0 }, { "word": "缩", "start": 215.19589903846153, "end": 215.31589903846154, "probability": 1.0 }, { "word": "后", "start": 215.31589903846154, "end": 215.45589903846155, "probability": 0.9970703125 }, { "word": "的", "start": 215.45589903846155, "end": 215.55589903846155, "probability": 0.98388671875 }, { "word": "内", "start": 215.55589903846155, "end": 215.67589903846152, "probability": 0.99853515625 }, { "word": "存", "start": 215.67589903846152, "end": 215.85589903846153, "probability": 0.99951171875 }, { "word": "里", "start": 215.85589903846153, "end": 216.01589903846153, "probability": 0.9765625 } ], "asr_metrics": { "avg_logprob": -0.11731879699406365, "compression_ratio": 1.234375, "no_speech_prob": 5.1019511532190975e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有预天穷算力HBM SSD 还有带宽如今都得适配这类需求这是因为我们大量使用智能体工具至少也是当下的模型服务模式导致的V4.1的优化思路从这里开始就很有意思了因为在V4版本中他们已经用压缩吸收注意力CSA将多个Token压缩为更少的KV条目之后再用吸收索引器判断这些压缩后的内存里哪些部分真正值得读取因此为了应对不断增长的上下文长度他们已经采用了" } ], "id": 102, "review_annotation": "", "row_id": 102, "bounded_row_id": 102, "parent_row_ids": [ 102 ], "parent_row_id": 102, "packet_id": 7, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 192.31589903846154, "packet_end": 221.50388461538464, "raw_text": "判断这些压缩后的内存里", "selected_text": "判断这些压缩后的内存里", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 102 ], "packet_id": 7, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 216.01589903846153, "end": 217.89589903846155, "text": "哪些部分真正值得读取", "raw_asr_text": "哪些部分真正值得读取", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 7, "words": [ { "word": "哪", "start": 216.01589903846153, "end": 216.27589903846155, "probability": 0.935546875 }, { "word": "些", "start": 216.27589903846155, "end": 216.45589903846155, "probability": 0.9970703125 }, { "word": "部分", "start": 216.45589903846155, "end": 216.75589903846154, "probability": 0.99609375 }, { "word": "真", "start": 216.75589903846154, "end": 217.05589903846155, "probability": 0.99267578125 }, { "word": "正", "start": 217.05589903846155, "end": 217.23589903846153, "probability": 1.0 }, { "word": "值", "start": 217.23589903846153, "end": 217.43589903846154, "probability": 1.0 }, { "word": "得", "start": 217.43589903846154, "end": 217.55589903846155, "probability": 1.0 }, { "word": "读", "start": 217.55589903846155, "end": 217.69589903846153, "probability": 0.99560546875 }, { "word": "取", "start": 217.69589903846153, "end": 217.89589903846155, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11731879699406365, "compression_ratio": 1.234375, "no_speech_prob": 5.1019511532190975e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有预天穷算力HBM SSD 还有带宽如今都得适配这类需求这是因为我们大量使用智能体工具至少也是当下的模型服务模式导致的V4.1的优化思路从这里开始就很有意思了因为在V4版本中他们已经用压缩吸收注意力CSA将多个Token压缩为更少的KV条目之后再用吸收索引器判断这些压缩后的内存里哪些部分真正值得读取因此为了应对不断增长的上下文长度他们已经采用了" } ], "id": 103, "review_annotation": "", "row_id": 103, "bounded_row_id": 103, "parent_row_ids": [ 103 ], "parent_row_id": 103, "packet_id": 7, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 192.31589903846154, "packet_end": 221.50388461538464, "raw_text": "哪些部分真正值得读取", "selected_text": "哪些部分真正值得读取", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 103 ], "packet_id": 7, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 217.89589903846155, "end": 220.65589903846154, "text": "因此为了应对不断增长的上下文长度", "raw_asr_text": "因此为了应对不断增长的上下文长度", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 7, "words": [ { "word": "因", "start": 217.89589903846155, "end": 218.31589903846154, "probability": 0.98828125 }, { "word": "此", "start": 218.31589903846154, "end": 218.47589903846153, "probability": 1.0 }, { "word": "为", "start": 218.47589903846153, "end": 218.63589903846153, "probability": 0.9970703125 }, { "word": "了", "start": 218.63589903846153, "end": 218.71589903846154, "probability": 0.99755859375 }, { "word": "应", "start": 218.71589903846154, "end": 218.83589903846155, "probability": 0.9970703125 }, { "word": "对", "start": 218.83589903846155, "end": 218.97589903846153, "probability": 0.9990234375 }, { "word": "不", "start": 218.97589903846153, "end": 219.21589903846154, "probability": 0.99755859375 }, { "word": "断", "start": 219.21589903846154, "end": 219.35589903846153, "probability": 0.9990234375 }, { "word": "增", "start": 219.35589903846153, "end": 219.63589903846153, "probability": 0.99853515625 }, { "word": "长", "start": 219.63589903846153, "end": 219.79589903846153, "probability": 0.9990234375 }, { "word": "的", "start": 219.79589903846153, "end": 219.91589903846153, "probability": 0.998046875 }, { "word": "上", "start": 219.91589903846153, "end": 220.05589903846155, "probability": 0.99658203125 }, { "word": "下", "start": 220.05589903846155, "end": 220.17589903846152, "probability": 0.9970703125 }, { "word": "文", "start": 220.17589903846152, "end": 220.31589903846154, "probability": 0.8935546875 }, { "word": "长", "start": 220.31589903846154, "end": 220.47589903846153, "probability": 0.9873046875 }, { "word": "度", "start": 220.47589903846153, "end": 220.65589903846154, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11731879699406365, "compression_ratio": 1.234375, "no_speech_prob": 5.1019511532190975e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有预天穷算力HBM SSD 还有带宽如今都得适配这类需求这是因为我们大量使用智能体工具至少也是当下的模型服务模式导致的V4.1的优化思路从这里开始就很有意思了因为在V4版本中他们已经用压缩吸收注意力CSA将多个Token压缩为更少的KV条目之后再用吸收索引器判断这些压缩后的内存里哪些部分真正值得读取因此为了应对不断增长的上下文长度他们已经采用了" } ], "id": 104, "review_annotation": "", "row_id": 104, "bounded_row_id": 104, "parent_row_ids": [ 104 ], "parent_row_id": 104, "packet_id": 7, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 192.31589903846154, "packet_end": 221.50388461538464, "raw_text": "因此为了应对不断增长的上下文长度", "selected_text": "因此为了应对不断增长的上下文长度", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 104 ], "packet_id": 7, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 220.70388461538462, "end": 222.86388461538462, "text": "他们已经采用了大量压缩手段", "raw_asr_text": "他们已经采用了大量压缩手段", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 8, "words": [ { "word": "他们", "start": 220.64388461538462, "end": 220.9038846153846, "probability": 0.66552734375 }, { "word": "已经", "start": 220.9038846153846, "end": 221.14388461538462, "probability": 0.99755859375 }, { "word": "采", "start": 221.14388461538462, "end": 221.3438846153846, "probability": 0.99951171875 }, { "word": "用", "start": 221.3438846153846, "end": 221.48388461538462, "probability": 1.0 }, { "word": "了", "start": 221.48388461538462, "end": 221.6238846153846, "probability": 0.998046875 }, { "word": "大", "start": 221.6238846153846, "end": 221.86388461538462, "probability": 0.9951171875 }, { "word": "量", "start": 221.86388461538462, "end": 222.1038846153846, "probability": 0.99951171875 }, { "word": "压", "start": 222.1038846153846, "end": 222.3838846153846, "probability": 0.9990234375 }, { "word": "缩", "start": 222.3838846153846, "end": 222.52388461538462, "probability": 0.9970703125 }, { "word": "手", "start": 222.52388461538462, "end": 222.6838846153846, "probability": 0.99853515625 }, { "word": "段", "start": 222.6838846153846, "end": 222.86388461538462, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.07360267178447927, "compression_ratio": 1.2877492877492878, "no_speech_prob": 2.1439724995353515e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "上下文长度他们已经采用了大量压缩手段但V4.1 Flash的论文指出KV缓存还会随着模型深度的增加而变大模型深度与模型规模高度相关因为如果想要更大的模型增加深度是最直接的选择之一所以在其他条件相同的情况下从12层增至120层可能意味着要存储约10倍的层专属KV状态即使你激进的压缩上下文本身无论剩下多少KV他们仍会沿网络深度成倍增加" } ], "id": 105, "review_annotation": "", "row_id": 105, "bounded_row_id": 105, "parent_row_ids": [ 105 ], "parent_row_id": 105, "packet_id": 8, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 219.9038846153846, "packet_end": 249.0918701923077, "raw_text": "他们已经采用了大量压缩手段", "selected_text": "他们已经采用了大量压缩手段", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 105 ], "packet_id": 8, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 222.86388461538462, "end": 225.08388461538462, "text": "但V4.1 Flash的论文指出", "raw_asr_text": "但V4.1 Flash的论文指出", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 8, "words": [ { "word": "但", "start": 222.86388461538462, "end": 223.26388461538463, "probability": 0.99072265625 }, { "word": "V", "start": 223.26388461538463, "end": 223.46388461538461, "probability": 0.46142578125 }, { "word": "4.", "start": 223.46388461538461, "end": 223.6638846153846, "probability": 0.98291015625 }, { "word": "1", "start": 223.76388461538463, "end": 223.8838846153846, "probability": 0.90673828125 }, { "word": " Flash", "start": 223.8838846153846, "end": 224.1238846153846, "probability": 0.32080078125 }, { "word": "的", "start": 224.1238846153846, "end": 224.36388461538462, "probability": 0.98193359375 }, { "word": "论", "start": 224.36388461538462, "end": 224.5038846153846, "probability": 0.99853515625 }, { "word": "文", "start": 224.5038846153846, "end": 224.64388461538462, "probability": 0.99951171875 }, { "word": "指", "start": 224.64388461538462, "end": 224.8238846153846, "probability": 0.9990234375 }, { "word": "出", "start": 224.8238846153846, "end": 225.08388461538462, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.07360267178447927, "compression_ratio": 1.2877492877492878, "no_speech_prob": 2.1439724995353515e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "上下文长度他们已经采用了大量压缩手段但V4.1 Flash的论文指出KV缓存还会随着模型深度的增加而变大模型深度与模型规模高度相关因为如果想要更大的模型增加深度是最直接的选择之一所以在其他条件相同的情况下从12层增至120层可能意味着要存储约10倍的层专属KV状态即使你激进的压缩上下文本身无论剩下多少KV他们仍会沿网络深度成倍增加" } ], "id": 106, "review_annotation": "", "row_id": 106, "bounded_row_id": 106, "parent_row_ids": [ 106 ], "parent_row_id": 106, "packet_id": 8, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 219.9038846153846, "packet_end": 249.0918701923077, "raw_text": "但V4.1 Flash的论文指出", "selected_text": "但V4.1 Flash的论文指出", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 106 ], "packet_id": 8, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 225.08388461538462, "end": 228.58388461538462, "text": "KV缓存还会随着模型深度的增加而变大", "raw_asr_text": "KV缓存还会随着模型深度的增加而变大", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 8, "words": [ { "word": "K", "start": 225.08388461538462, "end": 225.36388461538462, "probability": 0.9267578125 }, { "word": "V", "start": 225.36388461538462, "end": 225.52388461538462, "probability": 0.98095703125 }, { "word": "缓", "start": 225.52388461538462, "end": 225.74388461538462, "probability": 0.99658203125 }, { "word": "存", "start": 225.74388461538462, "end": 225.96388461538461, "probability": 1.0 }, { "word": "还", "start": 225.96388461538461, "end": 226.2238846153846, "probability": 0.998046875 }, { "word": "会", "start": 226.2238846153846, "end": 226.4038846153846, "probability": 0.99951171875 }, { "word": "随", "start": 226.4038846153846, "end": 226.58388461538462, "probability": 1.0 }, { "word": "着", "start": 226.58388461538462, "end": 226.7238846153846, "probability": 0.9990234375 }, { "word": "模", "start": 226.7238846153846, "end": 226.9038846153846, "probability": 0.99853515625 }, { "word": "型", "start": 226.9038846153846, "end": 227.1238846153846, "probability": 0.99951171875 }, { "word": "深", "start": 227.1238846153846, "end": 227.30388461538462, "probability": 0.99462890625 }, { "word": "度", "start": 227.30388461538462, "end": 227.5438846153846, "probability": 1.0 }, { "word": "的", "start": 227.5438846153846, "end": 227.6638846153846, "probability": 0.9970703125 }, { "word": "增", "start": 227.6638846153846, "end": 227.86388461538462, "probability": 1.0 }, { "word": "加", "start": 227.86388461538462, "end": 228.0638846153846, "probability": 0.99951171875 }, { "word": "而", "start": 228.0638846153846, "end": 228.24388461538462, "probability": 0.9990234375 }, { "word": "变", "start": 228.24388461538462, "end": 228.4038846153846, "probability": 1.0 }, { "word": "大", "start": 228.4038846153846, "end": 228.58388461538462, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.07360267178447927, "compression_ratio": 1.2877492877492878, "no_speech_prob": 2.1439724995353515e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "上下文长度他们已经采用了大量压缩手段但V4.1 Flash的论文指出KV缓存还会随着模型深度的增加而变大模型深度与模型规模高度相关因为如果想要更大的模型增加深度是最直接的选择之一所以在其他条件相同的情况下从12层增至120层可能意味着要存储约10倍的层专属KV状态即使你激进的压缩上下文本身无论剩下多少KV他们仍会沿网络深度成倍增加" } ], "id": 107, "review_annotation": "", "row_id": 107, "bounded_row_id": 107, "parent_row_ids": [ 107 ], "parent_row_id": 107, "packet_id": 8, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 219.9038846153846, "packet_end": 249.0918701923077, "raw_text": "KV缓存还会随着模型深度的增加而变大", "selected_text": "KV缓存还会随着模型深度的增加而变大", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 107 ], "packet_id": 8, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 228.58388461538462, "end": 231.24388461538462, "text": "模型深度与模型规模高度相关", "raw_asr_text": "模型深度与模型规模高度相关", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 8, "words": [ { "word": "模", "start": 228.58388461538462, "end": 228.9438846153846, "probability": 0.99560546875 }, { "word": "型", "start": 228.9438846153846, "end": 229.1238846153846, "probability": 1.0 }, { "word": "深", "start": 229.1238846153846, "end": 229.3438846153846, "probability": 0.9990234375 }, { "word": "度", "start": 229.3438846153846, "end": 229.6038846153846, "probability": 1.0 }, { "word": "与", "start": 229.6038846153846, "end": 229.8238846153846, "probability": 0.99853515625 }, { "word": "模", "start": 229.8238846153846, "end": 230.0038846153846, "probability": 0.99951171875 }, { "word": "型", "start": 230.0038846153846, "end": 230.1838846153846, "probability": 1.0 }, { "word": "规", "start": 230.1838846153846, "end": 230.36388461538462, "probability": 0.99951171875 }, { "word": "模", "start": 230.36388461538462, "end": 230.5038846153846, "probability": 0.99951171875 }, { "word": "高", "start": 230.5038846153846, "end": 230.7238846153846, "probability": 0.99853515625 }, { "word": "度", "start": 230.7238846153846, "end": 230.8838846153846, "probability": 1.0 }, { "word": "相", "start": 230.8838846153846, "end": 231.02388461538462, "probability": 1.0 }, { "word": "关", "start": 231.02388461538462, "end": 231.24388461538462, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.07360267178447927, "compression_ratio": 1.2877492877492878, "no_speech_prob": 2.1439724995353515e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "上下文长度他们已经采用了大量压缩手段但V4.1 Flash的论文指出KV缓存还会随着模型深度的增加而变大模型深度与模型规模高度相关因为如果想要更大的模型增加深度是最直接的选择之一所以在其他条件相同的情况下从12层增至120层可能意味着要存储约10倍的层专属KV状态即使你激进的压缩上下文本身无论剩下多少KV他们仍会沿网络深度成倍增加" } ], "id": 108, "review_annotation": "", "row_id": 108, "bounded_row_id": 108, "parent_row_ids": [ 108 ], "parent_row_id": 108, "packet_id": 8, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 219.9038846153846, "packet_end": 249.0918701923077, "raw_text": "模型深度与模型规模高度相关", "selected_text": "模型深度与模型规模高度相关", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 108 ], "packet_id": 8, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 231.24388461538462, "end": 233.2638846153846, "text": "因为如果想要更大的模型", "raw_asr_text": "因为如果想要更大的模型", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 8, "words": [ { "word": "因为", "start": 231.24388461538462, "end": 231.6838846153846, "probability": 0.9970703125 }, { "word": "如果", "start": 231.6838846153846, "end": 232.02388461538462, "probability": 0.9990234375 }, { "word": "想", "start": 232.02388461538462, "end": 232.2238846153846, "probability": 0.99951171875 }, { "word": "要", "start": 232.2238846153846, "end": 232.36388461538462, "probability": 0.99951171875 }, { "word": "更", "start": 232.36388461538462, "end": 232.54388461538463, "probability": 0.9990234375 }, { "word": "大的", "start": 232.54388461538463, "end": 232.8438846153846, "probability": 0.99658203125 }, { "word": "模", "start": 232.8438846153846, "end": 233.0638846153846, "probability": 0.99951171875 }, { "word": "型", "start": 233.0638846153846, "end": 233.2638846153846, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07360267178447927, "compression_ratio": 1.2877492877492878, "no_speech_prob": 2.1439724995353515e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "上下文长度他们已经采用了大量压缩手段但V4.1 Flash的论文指出KV缓存还会随着模型深度的增加而变大模型深度与模型规模高度相关因为如果想要更大的模型增加深度是最直接的选择之一所以在其他条件相同的情况下从12层增至120层可能意味着要存储约10倍的层专属KV状态即使你激进的压缩上下文本身无论剩下多少KV他们仍会沿网络深度成倍增加" } ], "id": 109, "review_annotation": "", "row_id": 109, "bounded_row_id": 109, "parent_row_ids": [ 109 ], "parent_row_id": 109, "packet_id": 8, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 219.9038846153846, "packet_end": 249.0918701923077, "raw_text": "因为如果想要更大的模型", "selected_text": "因为如果想要更大的模型", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 109 ], "packet_id": 8, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 233.2638846153846, "end": 235.24388461538462, "text": "增加深度是最直接的选择之一", "raw_asr_text": "增加深度是最直接的选择之一", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 8, "words": [ { "word": "增", "start": 233.2638846153846, "end": 233.4438846153846, "probability": 0.9970703125 }, { "word": "加", "start": 233.4438846153846, "end": 233.58388461538462, "probability": 0.9970703125 }, { "word": "深", "start": 233.58388461538462, "end": 233.7238846153846, "probability": 0.9990234375 }, { "word": "度", "start": 233.7238846153846, "end": 233.8838846153846, "probability": 1.0 }, { "word": "是", "start": 233.8838846153846, "end": 234.02388461538462, "probability": 0.998046875 }, { "word": "最", "start": 234.02388461538462, "end": 234.2238846153846, "probability": 0.99951171875 }, { "word": "直接", "start": 234.2238846153846, "end": 234.48388461538462, "probability": 0.9970703125 }, { "word": "的", "start": 234.48388461538462, "end": 234.64388461538462, "probability": 0.99951171875 }, { "word": "选", "start": 234.64388461538462, "end": 234.80388461538462, "probability": 0.99951171875 }, { "word": "择", "start": 234.80388461538462, "end": 234.96388461538461, "probability": 1.0 }, { "word": "之", "start": 234.96388461538461, "end": 235.08388461538462, "probability": 0.99951171875 }, { "word": "一", "start": 235.08388461538462, "end": 235.24388461538462, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.07360267178447927, "compression_ratio": 1.2877492877492878, "no_speech_prob": 2.1439724995353515e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "上下文长度他们已经采用了大量压缩手段但V4.1 Flash的论文指出KV缓存还会随着模型深度的增加而变大模型深度与模型规模高度相关因为如果想要更大的模型增加深度是最直接的选择之一所以在其他条件相同的情况下从12层增至120层可能意味着要存储约10倍的层专属KV状态即使你激进的压缩上下文本身无论剩下多少KV他们仍会沿网络深度成倍增加" } ], "id": 110, "review_annotation": "", "row_id": 110, "bounded_row_id": 110, "parent_row_ids": [ 110 ], "parent_row_id": 110, "packet_id": 8, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 219.9038846153846, "packet_end": 249.0918701923077, "raw_text": "增加深度是最直接的选择之一", "selected_text": "增加深度是最直接的选择之一", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 110 ], "packet_id": 8, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 235.24388461538462, "end": 237.6038846153846, "text": "所以在其他条件相同的情况下", "raw_asr_text": "所以在其他条件相同的情况下", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 8, "words": [ { "word": "所以", "start": 235.24388461538462, "end": 235.64388461538462, "probability": 0.9990234375 }, { "word": "在", "start": 235.64388461538462, "end": 235.9038846153846, "probability": 0.99658203125 }, { "word": "其", "start": 235.9038846153846, "end": 236.1638846153846, "probability": 0.9990234375 }, { "word": "他", "start": 236.1638846153846, "end": 236.3238846153846, "probability": 0.93017578125 }, { "word": "条", "start": 236.3238846153846, "end": 236.5038846153846, "probability": 0.99951171875 }, { "word": "件", "start": 236.5038846153846, "end": 236.64388461538462, "probability": 1.0 }, { "word": "相", "start": 236.64388461538462, "end": 236.80388461538462, "probability": 0.998046875 }, { "word": "同", "start": 236.80388461538462, "end": 237.02388461538462, "probability": 1.0 }, { "word": "的", "start": 237.02388461538462, "end": 237.14388461538462, "probability": 0.99853515625 }, { "word": "情况", "start": 237.14388461538462, "end": 237.3838846153846, "probability": 0.99951171875 }, { "word": "下", "start": 237.3838846153846, "end": 237.6038846153846, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.07360267178447927, "compression_ratio": 1.2877492877492878, "no_speech_prob": 2.1439724995353515e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "上下文长度他们已经采用了大量压缩手段但V4.1 Flash的论文指出KV缓存还会随着模型深度的增加而变大模型深度与模型规模高度相关因为如果想要更大的模型增加深度是最直接的选择之一所以在其他条件相同的情况下从12层增至120层可能意味着要存储约10倍的层专属KV状态即使你激进的压缩上下文本身无论剩下多少KV他们仍会沿网络深度成倍增加" } ], "id": 111, "review_annotation": "", "row_id": 111, "bounded_row_id": 111, "parent_row_ids": [ 111 ], "parent_row_id": 111, "packet_id": 8, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 219.9038846153846, "packet_end": 249.0918701923077, "raw_text": "所以在其他条件相同的情况下", "selected_text": "所以在其他条件相同的情况下", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 111 ], "packet_id": 8, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 237.6038846153846, "end": 239.8238846153846, "text": "从12层增至120层", "raw_asr_text": "从12层增至120层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 8, "words": [ { "word": "从", "start": 237.6038846153846, "end": 237.92388461538462, "probability": 0.99609375 }, { "word": "12", "start": 237.92388461538462, "end": 238.30388461538462, "probability": 0.99267578125 }, { "word": "层", "start": 238.30388461538462, "end": 238.5638846153846, "probability": 0.998046875 }, { "word": "增", "start": 238.5638846153846, "end": 238.8838846153846, "probability": 0.998046875 }, { "word": "至", "start": 238.8838846153846, "end": 239.04388461538463, "probability": 0.9970703125 }, { "word": "120", "start": 239.04388461538463, "end": 239.5038846153846, "probability": 0.998046875 }, { "word": "层", "start": 239.5038846153846, "end": 239.8238846153846, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.07360267178447927, "compression_ratio": 1.2877492877492878, "no_speech_prob": 2.1439724995353515e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "上下文长度他们已经采用了大量压缩手段但V4.1 Flash的论文指出KV缓存还会随着模型深度的增加而变大模型深度与模型规模高度相关因为如果想要更大的模型增加深度是最直接的选择之一所以在其他条件相同的情况下从12层增至120层可能意味着要存储约10倍的层专属KV状态即使你激进的压缩上下文本身无论剩下多少KV他们仍会沿网络深度成倍增加" } ], "id": 112, "review_annotation": "", "row_id": 112, "bounded_row_id": 112, "parent_row_ids": [ 112 ], "parent_row_id": 112, "packet_id": 8, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 219.9038846153846, "packet_end": 249.0918701923077, "raw_text": "从12层增至120层", "selected_text": "从12层增至120层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 112 ], "packet_id": 8, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 239.8238846153846, "end": 243.1038846153846, "text": "可能意味着要存储约10倍的层专属KV状态", "raw_asr_text": "可能意味着要存储约10倍的层专属KV状态", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 8, "words": [ { "word": "可能", "start": 239.8238846153846, "end": 240.1838846153846, "probability": 0.99609375 }, { "word": "意", "start": 240.1838846153846, "end": 240.3838846153846, "probability": 0.9990234375 }, { "word": "味", "start": 240.3838846153846, "end": 240.48388461538462, "probability": 0.9990234375 }, { "word": "着", "start": 240.48388461538462, "end": 240.6038846153846, "probability": 0.99853515625 }, { "word": "要", "start": 240.6038846153846, "end": 240.74388461538462, "probability": 0.99462890625 }, { "word": "存", "start": 240.74388461538462, "end": 241.0038846153846, "probability": 0.99853515625 }, { "word": "储", "start": 241.0038846153846, "end": 241.1638846153846, "probability": 0.998046875 }, { "word": "约", "start": 241.1638846153846, "end": 241.36388461538462, "probability": 0.99658203125 }, { "word": "10", "start": 241.36388461538462, "end": 241.64388461538462, "probability": 0.9609375 }, { "word": "倍", "start": 241.64388461538462, "end": 241.7838846153846, "probability": 0.99755859375 }, { "word": "的", "start": 241.7838846153846, "end": 241.9038846153846, "probability": 0.9921875 }, { "word": "层", "start": 241.9038846153846, "end": 242.08388461538462, "probability": 0.9970703125 }, { "word": "专", "start": 242.08388461538462, "end": 242.36388461538462, "probability": 0.99658203125 }, { "word": "属", "start": 242.36388461538462, "end": 242.5638846153846, "probability": 0.9990234375 }, { "word": "K", "start": 242.5638846153846, "end": 242.6838846153846, "probability": 0.97802734375 }, { "word": "V", "start": 242.6838846153846, "end": 242.7838846153846, "probability": 0.99853515625 }, { "word": "状", "start": 242.7838846153846, "end": 242.9438846153846, "probability": 0.9892578125 }, { "word": "态", "start": 242.9438846153846, "end": 243.1038846153846, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07360267178447927, "compression_ratio": 1.2877492877492878, "no_speech_prob": 2.1439724995353515e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "上下文长度他们已经采用了大量压缩手段但V4.1 Flash的论文指出KV缓存还会随着模型深度的增加而变大模型深度与模型规模高度相关因为如果想要更大的模型增加深度是最直接的选择之一所以在其他条件相同的情况下从12层增至120层可能意味着要存储约10倍的层专属KV状态即使你激进的压缩上下文本身无论剩下多少KV他们仍会沿网络深度成倍增加" } ], "id": 113, "review_annotation": "", "row_id": 113, "bounded_row_id": 113, "parent_row_ids": [ 113 ], "parent_row_id": 113, "packet_id": 8, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 219.9038846153846, "packet_end": 249.0918701923077, "raw_text": "可能意味着要存储约10倍的层专属KV状态", "selected_text": "可能意味着要存储约10倍的层专属KV状态", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 113 ], "packet_id": 8, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 243.1038846153846, "end": 245.2238846153846, "text": "即使你激进的压缩上下文本身", "raw_asr_text": "即使你激进的压缩上下文本身", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 8, "words": [ { "word": "即", "start": 243.1038846153846, "end": 243.30388461538462, "probability": 0.9931640625 }, { "word": "使", "start": 243.30388461538462, "end": 243.42388461538462, "probability": 0.99853515625 }, { "word": "你", "start": 243.42388461538462, "end": 243.6238846153846, "probability": 0.998046875 }, { "word": "激", "start": 243.6238846153846, "end": 243.80388461538462, "probability": 0.95556640625 }, { "word": "进", "start": 243.80388461538462, "end": 243.9438846153846, "probability": 0.99560546875 }, { "word": "的", "start": 243.9438846153846, "end": 244.04388461538463, "probability": 0.9619140625 }, { "word": "压", "start": 244.04388461538463, "end": 244.20388461538462, "probability": 1.0 }, { "word": "缩", "start": 244.20388461538462, "end": 244.3438846153846, "probability": 1.0 }, { "word": "上", "start": 244.3438846153846, "end": 244.52388461538462, "probability": 0.994140625 }, { "word": "下", "start": 244.52388461538462, "end": 244.6638846153846, "probability": 0.99951171875 }, { "word": "文", "start": 244.6638846153846, "end": 244.80388461538462, "probability": 0.99462890625 }, { "word": "本", "start": 244.80388461538462, "end": 245.0038846153846, "probability": 0.994140625 }, { "word": "身", "start": 245.0038846153846, "end": 245.2238846153846, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07360267178447927, "compression_ratio": 1.2877492877492878, "no_speech_prob": 2.1439724995353515e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "上下文长度他们已经采用了大量压缩手段但V4.1 Flash的论文指出KV缓存还会随着模型深度的增加而变大模型深度与模型规模高度相关因为如果想要更大的模型增加深度是最直接的选择之一所以在其他条件相同的情况下从12层增至120层可能意味着要存储约10倍的层专属KV状态即使你激进的压缩上下文本身无论剩下多少KV他们仍会沿网络深度成倍增加" } ], "id": 114, "review_annotation": "", "row_id": 114, "bounded_row_id": 114, "parent_row_ids": [ 114 ], "parent_row_id": 114, "packet_id": 8, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 219.9038846153846, "packet_end": 249.0918701923077, "raw_text": "即使你激进的压缩上下文本身", "selected_text": "即使你激进的压缩上下文本身", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 114 ], "packet_id": 8, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 245.2238846153846, "end": 246.70388461538462, "text": "无论剩下多少KV", "raw_asr_text": "无论剩下多少KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 8, "words": [ { "word": "无", "start": 245.2238846153846, "end": 245.5638846153846, "probability": 0.9951171875 }, { "word": "论", "start": 245.5638846153846, "end": 245.7238846153846, "probability": 1.0 }, { "word": "剩", "start": 245.7238846153846, "end": 245.8838846153846, "probability": 1.0 }, { "word": "下", "start": 245.8838846153846, "end": 246.04388461538463, "probability": 0.99853515625 }, { "word": "多少", "start": 246.04388461538463, "end": 246.2838846153846, "probability": 0.99853515625 }, { "word": "K", "start": 246.2838846153846, "end": 246.54388461538463, "probability": 0.99462890625 }, { "word": "V", "start": 246.54388461538463, "end": 246.70388461538462, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07360267178447927, "compression_ratio": 1.2877492877492878, "no_speech_prob": 2.1439724995353515e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "上下文长度他们已经采用了大量压缩手段但V4.1 Flash的论文指出KV缓存还会随着模型深度的增加而变大模型深度与模型规模高度相关因为如果想要更大的模型增加深度是最直接的选择之一所以在其他条件相同的情况下从12层增至120层可能意味着要存储约10倍的层专属KV状态即使你激进的压缩上下文本身无论剩下多少KV他们仍会沿网络深度成倍增加" } ], "id": 115, "review_annotation": "", "row_id": 115, "bounded_row_id": 115, "parent_row_ids": [ 115 ], "parent_row_id": 115, "packet_id": 8, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 219.9038846153846, "packet_end": 249.0918701923077, "raw_text": "无论剩下多少KV", "selected_text": "无论剩下多少KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 115 ], "packet_id": 8, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 246.70388461538462, "end": 248.2918701923077, "text": "他们仍会沿网络深度成倍增加", "raw_asr_text": "他们仍会沿网络深度成倍增加", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 8, "words": [ { "word": "他们", "start": 246.70388461538462, "end": 247.1038846153846, "probability": 0.4814453125 }, { "word": "仍", "start": 247.1038846153846, "end": 247.3438846153846, "probability": 0.99609375 }, { "word": "会", "start": 247.3438846153846, "end": 247.54388461538463, "probability": 0.998046875 }, { "word": "沿", "start": 247.54388461538463, "end": 247.7238846153846, "probability": 0.7724609375 }, { "word": "网", "start": 247.7238846153846, "end": 247.92388461538462, "probability": 0.99658203125 }, { "word": "络", "start": 247.92388461538462, "end": 248.04388461538463, "probability": 0.9970703125 }, { "word": "深", "start": 248.04388461538463, "end": 248.1838846153846, "probability": 0.9990234375 }, { "word": "度", "start": 248.1838846153846, "end": 248.3238846153846, "probability": 0.99951171875 }, { "word": "成", "start": 248.3238846153846, "end": 248.5038846153846, "probability": 0.85693359375 }, { "word": "倍", "start": 248.5038846153846, "end": 248.70388461538462, "probability": 0.998046875 }, { "word": "增", "start": 248.70388461538462, "end": 248.8838846153846, "probability": 0.99951171875 }, { "word": "加", "start": 248.8838846153846, "end": 249.0638846153846, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.07360267178447927, "compression_ratio": 1.2877492877492878, "no_speech_prob": 2.1439724995353515e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "上下文长度他们已经采用了大量压缩手段但V4.1 Flash的论文指出KV缓存还会随着模型深度的增加而变大模型深度与模型规模高度相关因为如果想要更大的模型增加深度是最直接的选择之一所以在其他条件相同的情况下从12层增至120层可能意味着要存储约10倍的层专属KV状态即使你激进的压缩上下文本身无论剩下多少KV他们仍会沿网络深度成倍增加" } ], "id": 116, "review_annotation": "", "row_id": 116, "bounded_row_id": 116, "parent_row_ids": [ 116 ], "parent_row_id": 116, "packet_id": 8, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 219.9038846153846, "packet_end": 249.0918701923077, "raw_text": "他们仍会沿网络深度成倍增加", "selected_text": "他们仍会沿网络深度成倍增加", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 116 ], "packet_id": 8, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 249.0718701923077, "end": 250.59187019230768, "text": "这就引出一个问题", "raw_asr_text": "这就引出一个问题", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "这", "start": 249.0718701923077, "end": 249.4918701923077, "probability": 0.96533203125 }, { "word": "就", "start": 249.4918701923077, "end": 249.63187019230767, "probability": 0.9931640625 }, { "word": "引", "start": 249.63187019230767, "end": 249.81187019230768, "probability": 0.9951171875 }, { "word": "出", "start": 249.81187019230768, "end": 250.0118701923077, "probability": 0.99658203125 }, { "word": "一个", "start": 250.0118701923077, "end": 250.1718701923077, "probability": 0.98681640625 }, { "word": "问题", "start": 250.1718701923077, "end": 250.59187019230768, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 117, "review_annotation": "", "row_id": 117, "bounded_row_id": 117, "parent_row_ids": [ 117 ], "parent_row_id": 117, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "这就引出一个问题", "selected_text": "这就引出一个问题", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 117 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 250.59187019230768, "end": 252.9918701923077, "text": "如果某一层已经生成了有用的", "raw_asr_text": "如果某一层已经生成了有用的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "如果", "start": 250.59187019230768, "end": 251.0518701923077, "probability": 0.96923828125 }, { "word": "某", "start": 251.0518701923077, "end": 251.31187019230768, "probability": 0.998046875 }, { "word": "一", "start": 251.31187019230768, "end": 251.43187019230768, "probability": 0.98828125 }, { "word": "层", "start": 251.43187019230768, "end": 251.63187019230767, "probability": 0.9990234375 }, { "word": "已经", "start": 251.63187019230767, "end": 251.91187019230767, "probability": 0.97998046875 }, { "word": "生", "start": 251.91187019230767, "end": 252.1718701923077, "probability": 0.95361328125 }, { "word": "成", "start": 252.1718701923077, "end": 252.3918701923077, "probability": 0.9951171875 }, { "word": "了", "start": 252.3918701923077, "end": 252.5118701923077, "probability": 0.990234375 }, { "word": "有", "start": 252.5118701923077, "end": 252.6718701923077, "probability": 0.99755859375 }, { "word": "用", "start": 252.6718701923077, "end": 252.8318701923077, "probability": 0.99951171875 }, { "word": "的", "start": 252.8318701923077, "end": 252.9918701923077, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 118, "review_annotation": "", "row_id": 118, "bounded_row_id": 118, "parent_row_ids": [ 118 ], "parent_row_id": 118, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "如果某一层已经生成了有用的", "selected_text": "如果某一层已经生成了有用的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 118 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 252.9918701923077, "end": 254.1118701923077, "text": "全局记忆表示", "raw_asr_text": "全局记忆表示", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "全", "start": 252.9918701923077, "end": 253.2718701923077, "probability": 0.92431640625 }, { "word": "局", "start": 253.2718701923077, "end": 253.47187019230768, "probability": 0.99853515625 }, { "word": "记", "start": 253.47187019230768, "end": 253.65187019230768, "probability": 0.89306640625 }, { "word": "忆", "start": 253.65187019230768, "end": 253.7718701923077, "probability": 0.99462890625 }, { "word": "表示", "start": 253.7718701923077, "end": 254.1118701923077, "probability": 0.94677734375 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 119, "review_annotation": "", "row_id": 119, "bounded_row_id": 119, "parent_row_ids": [ 119 ], "parent_row_id": 119, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "全局记忆表示", "selected_text": "全局记忆表示", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 119 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 254.1118701923077, "end": 256.8918701923077, "text": "下一层真的需要从头构建一个", "raw_asr_text": "下一层真的需要从头构建一个", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "下一", "start": 254.1118701923077, "end": 254.71187019230769, "probability": 0.9873046875 }, { "word": "层", "start": 254.71187019230769, "end": 255.0718701923077, "probability": 0.9990234375 }, { "word": "真的", "start": 255.0718701923077, "end": 255.3518701923077, "probability": 0.9912109375 }, { "word": "需要", "start": 255.3518701923077, "end": 255.71187019230769, "probability": 0.99072265625 }, { "word": "从", "start": 255.71187019230769, "end": 256.0118701923077, "probability": 0.97998046875 }, { "word": "头", "start": 256.0118701923077, "end": 256.2718701923077, "probability": 0.9970703125 }, { "word": "构", "start": 256.2718701923077, "end": 256.4918701923077, "probability": 0.99755859375 }, { "word": "建", "start": 256.4918701923077, "end": 256.6518701923077, "probability": 0.9951171875 }, { "word": "一个", "start": 256.6518701923077, "end": 256.8918701923077, "probability": 0.9921875 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 120, "review_annotation": "", "row_id": 120, "bounded_row_id": 120, "parent_row_ids": [ 120 ], "parent_row_id": 120, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "下一层真的需要从头构建一个", "selected_text": "下一层真的需要从头构建一个", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 120 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 256.8918701923077, "end": 258.05187019230766, "text": "全新的表示吗", "raw_asr_text": "全新的表示吗", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "全", "start": 256.8918701923077, "end": 257.1918701923077, "probability": 0.998046875 }, { "word": "新", "start": 257.1918701923077, "end": 257.3518701923077, "probability": 0.9951171875 }, { "word": "的", "start": 257.3518701923077, "end": 257.4518701923077, "probability": 0.99267578125 }, { "word": "表示", "start": 257.4518701923077, "end": 257.7718701923077, "probability": 0.9931640625 }, { "word": "吗", "start": 257.7718701923077, "end": 258.05187019230766, "probability": 0.98095703125 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 121, "review_annotation": "", "row_id": 121, "bounded_row_id": 121, "parent_row_ids": [ 121 ], "parent_row_id": 121, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "全新的表示吗", "selected_text": "全新的表示吗", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 121 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 258.05187019230766, "end": 259.11187019230766, "text": "Deepsy发现", "raw_asr_text": "Deepsy发现", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "De", "start": 258.05187019230766, "end": 258.4318701923077, "probability": 0.7578125 }, { "word": "ep", "start": 258.4318701923077, "end": 258.5318701923077, "probability": 0.93212890625 }, { "word": "sy", "start": 258.5318701923077, "end": 258.6318701923077, "probability": 0.5126953125 }, { "word": "发", "start": 258.6318701923077, "end": 258.8718701923077, "probability": 0.90185546875 }, { "word": "现", "start": 258.8718701923077, "end": 259.11187019230766, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 122, "review_annotation": "", "row_id": 122, "bounded_row_id": 122, "parent_row_ids": [ 122 ], "parent_row_id": 122, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "Deepsy发现", "selected_text": "Deepsy发现", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 122 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 259.11187019230766, "end": 260.5918701923077, "text": "至少对某些层来说", "raw_asr_text": "至少对某些层来说", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "至", "start": 259.11187019230766, "end": 259.3918701923077, "probability": 0.93994140625 }, { "word": "少", "start": 259.3918701923077, "end": 259.55187019230766, "probability": 1.0 }, { "word": "对", "start": 259.55187019230766, "end": 259.6918701923077, "probability": 0.99755859375 }, { "word": "某", "start": 259.6918701923077, "end": 259.8918701923077, "probability": 0.9990234375 }, { "word": "些", "start": 259.8918701923077, "end": 260.0118701923077, "probability": 0.98681640625 }, { "word": "层", "start": 260.0118701923077, "end": 260.17187019230767, "probability": 0.9990234375 }, { "word": "来", "start": 260.17187019230767, "end": 260.3318701923077, "probability": 0.9970703125 }, { "word": "说", "start": 260.3318701923077, "end": 260.5918701923077, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 123, "review_annotation": "", "row_id": 123, "bounded_row_id": 123, "parent_row_ids": [ 123 ], "parent_row_id": 123, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "至少对某些层来说", "selected_text": "至少对某些层来说", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 123 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 260.5918701923077, "end": 261.8918701923077, "text": "答案是否定的", "raw_asr_text": "答案是否定的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "答", "start": 260.5918701923077, "end": 260.9918701923077, "probability": 0.97802734375 }, { "word": "案", "start": 260.9918701923077, "end": 261.1518701923077, "probability": 0.99951171875 }, { "word": "是", "start": 261.1518701923077, "end": 261.3118701923077, "probability": 0.9990234375 }, { "word": "否", "start": 261.3118701923077, "end": 261.5118701923077, "probability": 1.0 }, { "word": "定", "start": 261.5118701923077, "end": 261.73187019230767, "probability": 0.99755859375 }, { "word": "的", "start": 261.73187019230767, "end": 261.8918701923077, "probability": 0.99560546875 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 124, "review_annotation": "", "row_id": 124, "bounded_row_id": 124, "parent_row_ids": [ 124 ], "parent_row_id": 124, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "答案是否定的", "selected_text": "答案是否定的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 124 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 261.8918701923077, "end": 263.23187019230767, "text": "这基本上就是压缩", "raw_asr_text": "这基本上就是压缩", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "这", "start": 261.8918701923077, "end": 262.05187019230766, "probability": 0.9736328125 }, { "word": "基本", "start": 262.05187019230766, "end": 262.3518701923077, "probability": 0.9951171875 }, { "word": "上", "start": 262.3518701923077, "end": 262.5118701923077, "probability": 0.99169921875 }, { "word": "就是", "start": 262.5118701923077, "end": 262.7118701923077, "probability": 0.98681640625 }, { "word": "压", "start": 262.7118701923077, "end": 263.0118701923077, "probability": 0.99072265625 }, { "word": "缩", "start": 263.0118701923077, "end": 263.23187019230767, "probability": 0.9970703125 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 125, "review_annotation": "", "row_id": 125, "bounded_row_id": 125, "parent_row_ids": [ 125 ], "parent_row_id": 125, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "这基本上就是压缩", "selected_text": "这基本上就是压缩", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 125 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 263.23187019230767, "end": 264.4318701923077, "text": "吸收注意力2", "raw_asr_text": "吸收注意力2", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "吸", "start": 263.23187019230767, "end": 263.5318701923077, "probability": 0.388427734375 }, { "word": "收", "start": 263.5318701923077, "end": 263.7118701923077, "probability": 0.6435546875 }, { "word": "注意", "start": 263.7118701923077, "end": 264.0118701923077, "probability": 0.98583984375 }, { "word": "力", "start": 264.0118701923077, "end": 264.1918701923077, "probability": 0.99755859375 }, { "word": "2", "start": 264.1918701923077, "end": 264.4318701923077, "probability": 0.9169921875 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 126, "review_annotation": "", "row_id": 126, "bounded_row_id": 126, "parent_row_ids": [ 126 ], "parent_row_id": 126, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "吸收注意力2", "selected_text": "吸收注意力2", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 126 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 264.4318701923077, "end": 266.2718701923077, "text": "CSAR背后的思路", "raw_asr_text": "CSAR背后的思路", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "CS", "start": 264.4318701923077, "end": 264.8518701923077, "probability": 0.31689453125 }, { "word": "AR", "start": 264.8518701923077, "end": 265.3718701923077, "probability": 0.90869140625 }, { "word": "背", "start": 265.3718701923077, "end": 265.67187019230767, "probability": 0.962890625 }, { "word": "后", "start": 265.67187019230767, "end": 265.8318701923077, "probability": 0.998046875 }, { "word": "的", "start": 265.8318701923077, "end": 265.9518701923077, "probability": 0.99755859375 }, { "word": "思", "start": 265.9518701923077, "end": 266.11187019230766, "probability": 0.99853515625 }, { "word": "路", "start": 266.11187019230766, "end": 266.2718701923077, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 127, "review_annotation": "", "row_id": 127, "bounded_row_id": 127, "parent_row_ids": [ 127 ], "parent_row_id": 127, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "CSAR背后的思路", "selected_text": "CSAR背后的思路", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 127 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 266.2718701923077, "end": 268.05187019230766, "text": "不再只沿Token纬度压缩", "raw_asr_text": "不再只沿Token纬度压缩", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "不", "start": 266.2718701923077, "end": 266.5718701923077, "probability": 0.9892578125 }, { "word": "再", "start": 266.5718701923077, "end": 266.7118701923077, "probability": 0.96923828125 }, { "word": "只", "start": 266.7118701923077, "end": 266.9118701923077, "probability": 0.78857421875 }, { "word": "沿", "start": 266.9118701923077, "end": 267.05187019230766, "probability": 0.8232421875 }, { "word": "T", "start": 267.05187019230766, "end": 267.2518701923077, "probability": 0.64453125 }, { "word": "oken", "start": 267.2518701923077, "end": 267.4118701923077, "probability": 0.97216796875 }, { "word": "纬", "start": 267.4118701923077, "end": 267.5318701923077, "probability": 0.7509765625 }, { "word": "度", "start": 267.5318701923077, "end": 267.67187019230767, "probability": 0.880859375 }, { "word": "压", "start": 267.67187019230767, "end": 267.8518701923077, "probability": 0.99853515625 }, { "word": "缩", "start": 267.8518701923077, "end": 268.05187019230766, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 128, "review_annotation": "", "row_id": 128, "bounded_row_id": 128, "parent_row_ids": [ 128 ], "parent_row_id": 128, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "不再只沿Token纬度压缩", "selected_text": "不再只沿Token纬度压缩", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 128 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 268.05187019230766, "end": 269.7118701923077, "text": "还开始沿层纬度压缩", "raw_asr_text": "还开始沿层纬度压缩", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "还", "start": 268.05187019230766, "end": 268.3518701923077, "probability": 0.9306640625 }, { "word": "开始", "start": 268.3518701923077, "end": 268.67187019230767, "probability": 0.99169921875 }, { "word": "沿", "start": 268.67187019230767, "end": 268.9318701923077, "probability": 0.900390625 }, { "word": "层", "start": 268.9318701923077, "end": 269.11187019230766, "probability": 0.97265625 }, { "word": "纬", "start": 269.11187019230766, "end": 269.2518701923077, "probability": 0.974609375 }, { "word": "度", "start": 269.2518701923077, "end": 269.3718701923077, "probability": 0.6884765625 }, { "word": "压", "start": 269.3718701923077, "end": 269.5318701923077, "probability": 0.99951171875 }, { "word": "缩", "start": 269.5318701923077, "end": 269.7118701923077, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 129, "review_annotation": "", "row_id": 129, "bounded_row_id": 129, "parent_row_ids": [ 129 ], "parent_row_id": 129, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "还开始沿层纬度压缩", "selected_text": "还开始沿层纬度压缩", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 129 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 269.7118701923077, "end": 271.2718701923077, "text": "他们的做法是将注意力层", "raw_asr_text": "他们的做法是将注意力层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "他们", "start": 269.7118701923077, "end": 270.05187019230766, "probability": 0.62158203125 }, { "word": "的", "start": 270.05187019230766, "end": 270.1918701923077, "probability": 0.9970703125 }, { "word": "做", "start": 270.1918701923077, "end": 270.3118701923077, "probability": 0.98388671875 }, { "word": "法", "start": 270.3118701923077, "end": 270.4718701923077, "probability": 0.99951171875 }, { "word": "是", "start": 270.4718701923077, "end": 270.61187019230766, "probability": 0.99658203125 }, { "word": "将", "start": 270.61187019230766, "end": 270.73187019230767, "probability": 0.98876953125 }, { "word": "注意", "start": 270.73187019230767, "end": 270.9518701923077, "probability": 0.98486328125 }, { "word": "力", "start": 270.9518701923077, "end": 271.11187019230766, "probability": 0.99462890625 }, { "word": "层", "start": 271.11187019230766, "end": 271.2718701923077, "probability": 0.966796875 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 130, "review_annotation": "", "row_id": 130, "bounded_row_id": 130, "parent_row_ids": [ 130 ], "parent_row_id": 130, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "他们的做法是将注意力层", "selected_text": "他们的做法是将注意力层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 130 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 271.2718701923077, "end": 272.61187019230766, "text": "分成三种不同模式", "raw_asr_text": "分成三种不同模式", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "分", "start": 271.2718701923077, "end": 271.4318701923077, "probability": 0.99609375 }, { "word": "成", "start": 271.4318701923077, "end": 271.5718701923077, "probability": 0.98876953125 }, { "word": "三", "start": 271.5718701923077, "end": 271.7118701923077, "probability": 0.9482421875 }, { "word": "种", "start": 271.7118701923077, "end": 271.8718701923077, "probability": 0.99853515625 }, { "word": "不同", "start": 271.8718701923077, "end": 272.11187019230766, "probability": 0.99365234375 }, { "word": "模", "start": 272.11187019230766, "end": 272.3718701923077, "probability": 0.9931640625 }, { "word": "式", "start": 272.3718701923077, "end": 272.61187019230766, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 131, "review_annotation": "", "row_id": 131, "bounded_row_id": 131, "parent_row_ids": [ 131 ], "parent_row_id": 131, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "分成三种不同模式", "selected_text": "分成三种不同模式", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 131 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 272.61187019230766, "end": 273.8118701923077, "text": "完整层模式", "raw_asr_text": "完整层模式", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "完", "start": 272.61187019230766, "end": 273.0718701923077, "probability": 0.96142578125 }, { "word": "整", "start": 273.0718701923077, "end": 273.2718701923077, "probability": 0.99951171875 }, { "word": "层", "start": 273.2718701923077, "end": 273.4118701923077, "probability": 0.9990234375 }, { "word": "模", "start": 273.4118701923077, "end": 273.5918701923077, "probability": 0.99560546875 }, { "word": "式", "start": 273.5918701923077, "end": 273.8118701923077, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 132, "review_annotation": "", "row_id": 132, "bounded_row_id": 132, "parent_row_ids": [ 132 ], "parent_row_id": 132, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "完整层模式", "selected_text": "完整层模式", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 132 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 273.8118701923077, "end": 275.11187019230766, "text": "重索引模式", "raw_asr_text": "重索引模式", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "重", "start": 273.8118701923077, "end": 274.3118701923077, "probability": 0.93115234375 }, { "word": "索", "start": 274.3118701923077, "end": 274.5118701923077, "probability": 0.8662109375 }, { "word": "引", "start": 274.5118701923077, "end": 274.7118701923077, "probability": 0.98779296875 }, { "word": "模", "start": 274.7118701923077, "end": 274.8718701923077, "probability": 0.99267578125 }, { "word": "式", "start": 274.8718701923077, "end": 275.11187019230766, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 133, "review_annotation": "", "row_id": 133, "bounded_row_id": 133, "parent_row_ids": [ 133 ], "parent_row_id": 133, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "重索引模式", "selected_text": "重索引模式", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 133 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 275.11187019230766, "end": 275.8798557692308, "text": "和复用模式", "raw_asr_text": "和复用模式", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 9, "words": [ { "word": "和", "start": 275.11187019230766, "end": 275.4118701923077, "probability": 0.94287109375 }, { "word": "复", "start": 275.4118701923077, "end": 275.5918701923077, "probability": 0.86083984375 }, { "word": "用", "start": 275.5918701923077, "end": 275.73187019230767, "probability": 0.98974609375 }, { "word": "模", "start": 275.73187019230767, "end": 275.8918701923077, "probability": 0.9990234375 }, { "word": "式", "start": 275.8918701923077, "end": 276.11187019230766, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1787575656927905, "compression_ratio": 1.3362573099415205, "no_speech_prob": 3.703489251383196e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "会以沿网络深度成倍增加这就引出一个问题如果某一层已经生成了有用的全局记忆表示下一层真的需要从头构建一个全新的表示吗Deepsy发现至少对某些层来说答案是否定的这基本上就是压缩吸收注意力2CSAR背后的思路不再只沿Token纬度压缩还开始沿层纬度压缩他们的做法是将注意力层分成三种不同模式完整层模式重索引模式和复用模式完整层模式" } ], "id": 134, "review_annotation": "", "row_id": 134, "bounded_row_id": 134, "parent_row_ids": [ 134 ], "parent_row_id": 134, "packet_id": 9, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 247.4918701923077, "packet_end": 276.6798557692308, "raw_text": "和复用模式", "selected_text": "和复用模式", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 134 ], "packet_id": 9, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 276.09985576923077, "end": 278.0198557692308, "text": "完整层会正常执行所有操作", "raw_asr_text": "完整层会正常执行所有操作", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 10, "words": [ { "word": "完", "start": 276.09985576923077, "end": 276.35985576923076, "probability": 0.9677734375 }, { "word": "整", "start": 276.35985576923076, "end": 276.5198557692308, "probability": 1.0 }, { "word": "层", "start": 276.5198557692308, "end": 276.6998557692308, "probability": 0.9990234375 }, { "word": "会", "start": 276.6998557692308, "end": 276.85985576923076, "probability": 0.99658203125 }, { "word": "正", "start": 276.85985576923076, "end": 276.9998557692308, "probability": 0.9970703125 }, { "word": "常", "start": 276.9998557692308, "end": 277.1598557692308, "probability": 1.0 }, { "word": "执", "start": 277.1598557692308, "end": 277.2998557692308, "probability": 0.9990234375 }, { "word": "行", "start": 277.2998557692308, "end": 277.4398557692308, "probability": 1.0 }, { "word": "所有", "start": 277.4398557692308, "end": 277.6998557692308, "probability": 0.99853515625 }, { "word": "操", "start": 277.6998557692308, "end": 277.8798557692308, "probability": 0.99951171875 }, { "word": "作", "start": 277.8798557692308, "end": 278.0198557692308, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09651627260095932, "compression_ratio": 1.4588607594936709, "no_speech_prob": 2.7198640908943084e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和复用模式完整层会正常执行所有操作它会生成全局KV生成锁引起K运行吸收锁引起并选择自己的TopK记忆条目如果需要快速回顾一下吸收注意力吸收锁引起基本上就是KV缓存的一个低成本搜索引擎它会给缓存的全局KV条目打分选出最相关的TopK条目而注意力只读取这些被选中的条目接下来是重锁引模式该层会使用上一层已经生成的全局KV但它对哪些信息" } ], "id": 135, "review_annotation": "", "row_id": 135, "bounded_row_id": 135, "parent_row_ids": [ 135 ], "parent_row_id": 135, "packet_id": 10, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 275.0798557692308, "packet_end": 304.2678413461539, "raw_text": "完整层会正常执行所有操作", "selected_text": "完整层会正常执行所有操作", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 135 ], "packet_id": 10, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 278.0198557692308, "end": 280.0198557692308, "text": "它会生成全局KV", "raw_asr_text": "它会生成全局KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 10, "words": [ { "word": "它", "start": 278.0198557692308, "end": 278.3798557692308, "probability": 0.96533203125 }, { "word": "会", "start": 278.3798557692308, "end": 278.53985576923077, "probability": 0.99951171875 }, { "word": "生", "start": 278.53985576923077, "end": 278.6998557692308, "probability": 0.99560546875 }, { "word": "成", "start": 278.6998557692308, "end": 278.9398557692308, "probability": 0.99365234375 }, { "word": "全", "start": 278.9398557692308, "end": 279.2998557692308, "probability": 0.994140625 }, { "word": "局", "start": 279.2998557692308, "end": 279.53985576923077, "probability": 0.99658203125 }, { "word": "K", "start": 279.53985576923077, "end": 279.7998557692308, "probability": 0.82763671875 }, { "word": "V", "start": 279.7998557692308, "end": 280.0198557692308, "probability": 0.99658203125 } ], "asr_metrics": { "avg_logprob": -0.09651627260095932, "compression_ratio": 1.4588607594936709, "no_speech_prob": 2.7198640908943084e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和复用模式完整层会正常执行所有操作它会生成全局KV生成锁引起K运行吸收锁引起并选择自己的TopK记忆条目如果需要快速回顾一下吸收注意力吸收锁引起基本上就是KV缓存的一个低成本搜索引擎它会给缓存的全局KV条目打分选出最相关的TopK条目而注意力只读取这些被选中的条目接下来是重锁引模式该层会使用上一层已经生成的全局KV但它对哪些信息" } ], "id": 136, "review_annotation": "", "row_id": 136, "bounded_row_id": 136, "parent_row_ids": [ 136 ], "parent_row_id": 136, "packet_id": 10, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 275.0798557692308, "packet_end": 304.2678413461539, "raw_text": "它会生成全局KV", "selected_text": "它会生成全局KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 136 ], "packet_id": 10, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 280.0198557692308, "end": 281.09985576923077, "text": "生成锁引起K", "raw_asr_text": "生成锁引起K", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 10, "words": [ { "word": "生", "start": 280.0198557692308, "end": 280.2798557692308, "probability": 0.97509765625 }, { "word": "成", "start": 280.2798557692308, "end": 280.5198557692308, "probability": 0.9990234375 }, { "word": "锁", "start": 280.5198557692308, "end": 280.6798557692308, "probability": 0.78271484375 }, { "word": "引", "start": 280.6798557692308, "end": 280.7998557692308, "probability": 0.84033203125 }, { "word": "起", "start": 280.7998557692308, "end": 280.91985576923076, "probability": 0.359619140625 }, { "word": "K", "start": 280.91985576923076, "end": 281.09985576923077, "probability": 0.84423828125 } ], "asr_metrics": { "avg_logprob": -0.09651627260095932, "compression_ratio": 1.4588607594936709, "no_speech_prob": 2.7198640908943084e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和复用模式完整层会正常执行所有操作它会生成全局KV生成锁引起K运行吸收锁引起并选择自己的TopK记忆条目如果需要快速回顾一下吸收注意力吸收锁引起基本上就是KV缓存的一个低成本搜索引擎它会给缓存的全局KV条目打分选出最相关的TopK条目而注意力只读取这些被选中的条目接下来是重锁引模式该层会使用上一层已经生成的全局KV但它对哪些信息" } ], "id": 137, "review_annotation": "", "row_id": 137, "bounded_row_id": 137, "parent_row_ids": [ 137 ], "parent_row_id": 137, "packet_id": 10, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 275.0798557692308, "packet_end": 304.2678413461539, "raw_text": "生成锁引起K", "selected_text": "生成锁引起K", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 137 ], "packet_id": 10, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 281.09985576923077, "end": 282.47985576923077, "text": "运行吸收锁引起", "raw_asr_text": "运行吸收锁引起", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 10, "words": [ { "word": "运", "start": 281.09985576923077, "end": 281.41985576923076, "probability": 0.9482421875 }, { "word": "行", "start": 281.41985576923076, "end": 281.6198557692308, "probability": 0.99951171875 }, { "word": "吸", "start": 281.6198557692308, "end": 281.85985576923076, "probability": 0.452880859375 }, { "word": "收", "start": 281.85985576923076, "end": 282.03985576923077, "probability": 0.71337890625 }, { "word": "锁", "start": 282.03985576923077, "end": 282.2198557692308, "probability": 0.9716796875 }, { "word": "引", "start": 282.2198557692308, "end": 282.3198557692308, "probability": 0.99951171875 }, { "word": "起", "start": 282.3198557692308, "end": 282.47985576923077, "probability": 0.96435546875 } ], "asr_metrics": { "avg_logprob": -0.09651627260095932, "compression_ratio": 1.4588607594936709, "no_speech_prob": 2.7198640908943084e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和复用模式完整层会正常执行所有操作它会生成全局KV生成锁引起K运行吸收锁引起并选择自己的TopK记忆条目如果需要快速回顾一下吸收注意力吸收锁引起基本上就是KV缓存的一个低成本搜索引擎它会给缓存的全局KV条目打分选出最相关的TopK条目而注意力只读取这些被选中的条目接下来是重锁引模式该层会使用上一层已经生成的全局KV但它对哪些信息" } ], "id": 138, "review_annotation": "", "row_id": 138, "bounded_row_id": 138, "parent_row_ids": [ 138 ], "parent_row_id": 138, "packet_id": 10, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 275.0798557692308, "packet_end": 304.2678413461539, "raw_text": "运行吸收锁引起", "selected_text": "运行吸收锁引起", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 138 ], "packet_id": 10, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 282.47985576923077, "end": 284.6598557692308, "text": "并选择自己的TopK记忆条目", "raw_asr_text": "并选择自己的TopK记忆条目", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 10, "words": [ { "word": "并", "start": 282.47985576923077, "end": 282.7198557692308, "probability": 0.99755859375 }, { "word": "选", "start": 282.7198557692308, "end": 282.91985576923076, "probability": 1.0 }, { "word": "择", "start": 282.91985576923076, "end": 283.09985576923077, "probability": 1.0 }, { "word": "自己", "start": 283.09985576923077, "end": 283.3798557692308, "probability": 0.9990234375 }, { "word": "的", "start": 283.3798557692308, "end": 283.5798557692308, "probability": 0.99853515625 }, { "word": "Top", "start": 283.5798557692308, "end": 283.8398557692308, "probability": 0.8994140625 }, { "word": "K", "start": 283.8398557692308, "end": 284.03985576923077, "probability": 0.76806640625 }, { "word": "记", "start": 284.03985576923077, "end": 284.2598557692308, "probability": 0.8994140625 }, { "word": "忆", "start": 284.2598557692308, "end": 284.3398557692308, "probability": 0.9970703125 }, { "word": "条", "start": 284.3398557692308, "end": 284.4598557692308, "probability": 0.99755859375 }, { "word": "目", "start": 284.4598557692308, "end": 284.6598557692308, "probability": 0.94775390625 } ], "asr_metrics": { "avg_logprob": -0.09651627260095932, "compression_ratio": 1.4588607594936709, "no_speech_prob": 2.7198640908943084e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和复用模式完整层会正常执行所有操作它会生成全局KV生成锁引起K运行吸收锁引起并选择自己的TopK记忆条目如果需要快速回顾一下吸收注意力吸收锁引起基本上就是KV缓存的一个低成本搜索引擎它会给缓存的全局KV条目打分选出最相关的TopK条目而注意力只读取这些被选中的条目接下来是重锁引模式该层会使用上一层已经生成的全局KV但它对哪些信息" } ], "id": 139, "review_annotation": "", "row_id": 139, "bounded_row_id": 139, "parent_row_ids": [ 139 ], "parent_row_id": 139, "packet_id": 10, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 275.0798557692308, "packet_end": 304.2678413461539, "raw_text": "并选择自己的TopK记忆条目", "selected_text": "并选择自己的TopK记忆条目", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 139 ], "packet_id": 10, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 284.6598557692308, "end": 287.2998557692308, "text": "如果需要快速回顾一下吸收注意力", "raw_asr_text": "如果需要快速回顾一下吸收注意力", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 10, "words": [ { "word": "如果", "start": 284.6598557692308, "end": 285.1198557692308, "probability": 0.99951171875 }, { "word": "需要", "start": 285.1198557692308, "end": 285.4398557692308, "probability": 0.99951171875 }, { "word": "快", "start": 285.4398557692308, "end": 285.6598557692308, "probability": 0.99951171875 }, { "word": "速", "start": 285.6598557692308, "end": 285.8398557692308, "probability": 1.0 }, { "word": "回", "start": 285.8398557692308, "end": 286.0198557692308, "probability": 0.99951171875 }, { "word": "顾", "start": 286.0198557692308, "end": 286.1998557692308, "probability": 0.9990234375 }, { "word": "一下", "start": 286.1998557692308, "end": 286.4398557692308, "probability": 0.99755859375 }, { "word": "吸", "start": 286.4398557692308, "end": 286.6998557692308, "probability": 0.9990234375 }, { "word": "收", "start": 286.6998557692308, "end": 286.8798557692308, "probability": 0.998046875 }, { "word": "注意", "start": 286.8798557692308, "end": 287.1198557692308, "probability": 0.998046875 }, { "word": "力", "start": 287.1198557692308, "end": 287.2998557692308, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.09651627260095932, "compression_ratio": 1.4588607594936709, "no_speech_prob": 2.7198640908943084e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和复用模式完整层会正常执行所有操作它会生成全局KV生成锁引起K运行吸收锁引起并选择自己的TopK记忆条目如果需要快速回顾一下吸收注意力吸收锁引起基本上就是KV缓存的一个低成本搜索引擎它会给缓存的全局KV条目打分选出最相关的TopK条目而注意力只读取这些被选中的条目接下来是重锁引模式该层会使用上一层已经生成的全局KV但它对哪些信息" } ], "id": 140, "review_annotation": "", "row_id": 140, "bounded_row_id": 140, "parent_row_ids": [ 140 ], "parent_row_id": 140, "packet_id": 10, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 275.0798557692308, "packet_end": 304.2678413461539, "raw_text": "如果需要快速回顾一下吸收注意力", "selected_text": "如果需要快速回顾一下吸收注意力", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 140 ], "packet_id": 10, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 287.2998557692308, "end": 291.0598557692308, "text": "吸收锁引起基本上就是KV缓存的一个低成本搜索引擎", "raw_asr_text": "吸收锁引起基本上就是KV缓存的一个低成本搜索引擎", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 10, "words": [ { "word": "吸", "start": 287.2998557692308, "end": 287.6998557692308, "probability": 0.9990234375 }, { "word": "收", "start": 287.6998557692308, "end": 287.8998557692308, "probability": 0.99951171875 }, { "word": "锁", "start": 287.8998557692308, "end": 288.09985576923077, "probability": 1.0 }, { "word": "引", "start": 288.09985576923077, "end": 288.2398557692308, "probability": 1.0 }, { "word": "起", "start": 288.2398557692308, "end": 288.3998557692308, "probability": 0.998046875 }, { "word": "基本", "start": 288.3998557692308, "end": 288.6798557692308, "probability": 0.99853515625 }, { "word": "上", "start": 288.6798557692308, "end": 288.85985576923076, "probability": 0.998046875 }, { "word": "就是", "start": 288.85985576923076, "end": 289.0598557692308, "probability": 0.9990234375 }, { "word": "K", "start": 289.0598557692308, "end": 289.2598557692308, "probability": 0.99658203125 }, { "word": "V", "start": 289.2598557692308, "end": 289.3998557692308, "probability": 0.99951171875 }, { "word": "缓", "start": 289.3998557692308, "end": 289.59985576923077, "probability": 0.998046875 }, { "word": "存", "start": 289.59985576923077, "end": 289.7198557692308, "probability": 1.0 }, { "word": "的", "start": 289.7198557692308, "end": 289.8398557692308, "probability": 0.99755859375 }, { "word": "一个", "start": 289.8398557692308, "end": 290.0198557692308, "probability": 0.97998046875 }, { "word": "低", "start": 290.0198557692308, "end": 290.2998557692308, "probability": 0.99853515625 }, { "word": "成", "start": 290.2998557692308, "end": 290.3998557692308, "probability": 0.994140625 }, { "word": "本", "start": 290.3998557692308, "end": 290.5198557692308, "probability": 0.99560546875 }, { "word": "搜", "start": 290.5198557692308, "end": 290.6798557692308, "probability": 0.9892578125 }, { "word": "索", "start": 290.6798557692308, "end": 290.7598557692308, "probability": 0.9970703125 }, { "word": "引", "start": 290.7598557692308, "end": 290.8798557692308, "probability": 1.0 }, { "word": "擎", "start": 290.8798557692308, "end": 291.0598557692308, "probability": 0.99609375 } ], "asr_metrics": { "avg_logprob": -0.09651627260095932, "compression_ratio": 1.4588607594936709, "no_speech_prob": 2.7198640908943084e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和复用模式完整层会正常执行所有操作它会生成全局KV生成锁引起K运行吸收锁引起并选择自己的TopK记忆条目如果需要快速回顾一下吸收注意力吸收锁引起基本上就是KV缓存的一个低成本搜索引擎它会给缓存的全局KV条目打分选出最相关的TopK条目而注意力只读取这些被选中的条目接下来是重锁引模式该层会使用上一层已经生成的全局KV但它对哪些信息" } ], "id": 141, "review_annotation": "", "row_id": 141, "bounded_row_id": 141, "parent_row_ids": [ 141 ], "parent_row_id": 141, "packet_id": 10, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 275.0798557692308, "packet_end": 304.2678413461539, "raw_text": "吸收锁引起基本上就是KV缓存的一个低成本搜索引擎", "selected_text": "吸收锁引起基本上就是KV缓存的一个低成本搜索引擎", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 141 ], "packet_id": 10, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 291.0598557692308, "end": 293.2998557692308, "text": "它会给缓存的全局KV条目打分", "raw_asr_text": "它会给缓存的全局KV条目打分", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 10, "words": [ { "word": "它", "start": 291.0598557692308, "end": 291.2398557692308, "probability": 0.98046875 }, { "word": "会", "start": 291.2398557692308, "end": 291.3798557692308, "probability": 1.0 }, { "word": "给", "start": 291.3798557692308, "end": 291.4998557692308, "probability": 1.0 }, { "word": "缓", "start": 291.4998557692308, "end": 291.7398557692308, "probability": 1.0 }, { "word": "存", "start": 291.7398557692308, "end": 291.8798557692308, "probability": 1.0 }, { "word": "的", "start": 291.8798557692308, "end": 292.0198557692308, "probability": 0.99951171875 }, { "word": "全", "start": 292.0198557692308, "end": 292.2198557692308, "probability": 0.99951171875 }, { "word": "局", "start": 292.2198557692308, "end": 292.35985576923076, "probability": 1.0 }, { "word": "K", "start": 292.35985576923076, "end": 292.5198557692308, "probability": 0.99951171875 }, { "word": "V", "start": 292.5198557692308, "end": 292.6598557692308, "probability": 1.0 }, { "word": "条", "start": 292.6598557692308, "end": 292.7998557692308, "probability": 0.9990234375 }, { "word": "目", "start": 292.7998557692308, "end": 292.97985576923077, "probability": 0.99951171875 }, { "word": "打", "start": 292.97985576923077, "end": 293.09985576923077, "probability": 0.99609375 }, { "word": "分", "start": 293.09985576923077, "end": 293.2998557692308, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09651627260095932, "compression_ratio": 1.4588607594936709, "no_speech_prob": 2.7198640908943084e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和复用模式完整层会正常执行所有操作它会生成全局KV生成锁引起K运行吸收锁引起并选择自己的TopK记忆条目如果需要快速回顾一下吸收注意力吸收锁引起基本上就是KV缓存的一个低成本搜索引擎它会给缓存的全局KV条目打分选出最相关的TopK条目而注意力只读取这些被选中的条目接下来是重锁引模式该层会使用上一层已经生成的全局KV但它对哪些信息" } ], "id": 142, "review_annotation": "", "row_id": 142, "bounded_row_id": 142, "parent_row_ids": [ 142 ], "parent_row_id": 142, "packet_id": 10, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 275.0798557692308, "packet_end": 304.2678413461539, "raw_text": "它会给缓存的全局KV条目打分", "selected_text": "它会给缓存的全局KV条目打分", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 142 ], "packet_id": 10, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 293.2998557692308, "end": 295.3798557692308, "text": "选出最相关的TopK条目", "raw_asr_text": "选出最相关的TopK条目", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 10, "words": [ { "word": "选", "start": 293.2998557692308, "end": 293.6198557692308, "probability": 0.99853515625 }, { "word": "出", "start": 293.6198557692308, "end": 293.7598557692308, "probability": 0.998046875 }, { "word": "最", "start": 293.7598557692308, "end": 293.9398557692308, "probability": 1.0 }, { "word": "相", "start": 293.9398557692308, "end": 294.1598557692308, "probability": 0.99951171875 }, { "word": "关", "start": 294.1598557692308, "end": 294.3398557692308, "probability": 0.99951171875 }, { "word": "的", "start": 294.3398557692308, "end": 294.4998557692308, "probability": 0.99951171875 }, { "word": "Top", "start": 294.4998557692308, "end": 294.7398557692308, "probability": 0.9931640625 }, { "word": "K", "start": 294.7398557692308, "end": 294.91985576923076, "probability": 0.99853515625 }, { "word": "条", "start": 294.91985576923076, "end": 295.1598557692308, "probability": 0.89208984375 }, { "word": "目", "start": 295.1598557692308, "end": 295.3798557692308, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09651627260095932, "compression_ratio": 1.4588607594936709, "no_speech_prob": 2.7198640908943084e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和复用模式完整层会正常执行所有操作它会生成全局KV生成锁引起K运行吸收锁引起并选择自己的TopK记忆条目如果需要快速回顾一下吸收注意力吸收锁引起基本上就是KV缓存的一个低成本搜索引擎它会给缓存的全局KV条目打分选出最相关的TopK条目而注意力只读取这些被选中的条目接下来是重锁引模式该层会使用上一层已经生成的全局KV但它对哪些信息" } ], "id": 143, "review_annotation": "", "row_id": 143, "bounded_row_id": 143, "parent_row_ids": [ 143 ], "parent_row_id": 143, "packet_id": 10, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 275.0798557692308, "packet_end": 304.2678413461539, "raw_text": "选出最相关的TopK条目", "selected_text": "选出最相关的TopK条目", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 143 ], "packet_id": 10, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 295.3798557692308, "end": 297.85985576923076, "text": "而注意力只读取这些被选中的条目", "raw_asr_text": "而注意力只读取这些被选中的条目", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 10, "words": [ { "word": "而", "start": 295.3798557692308, "end": 295.59985576923077, "probability": 0.99609375 }, { "word": "注意", "start": 295.59985576923077, "end": 295.7998557692308, "probability": 0.99267578125 }, { "word": "力", "start": 295.7998557692308, "end": 295.9598557692308, "probability": 0.99951171875 }, { "word": "只", "start": 295.9598557692308, "end": 296.2398557692308, "probability": 0.99853515625 }, { "word": "读", "start": 296.2398557692308, "end": 296.3798557692308, "probability": 0.9931640625 }, { "word": "取", "start": 296.3798557692308, "end": 296.53985576923077, "probability": 1.0 }, { "word": "这", "start": 296.53985576923077, "end": 296.6998557692308, "probability": 0.99853515625 }, { "word": "些", "start": 296.6998557692308, "end": 296.8198557692308, "probability": 1.0 }, { "word": "被", "start": 296.8198557692308, "end": 297.0198557692308, "probability": 0.99951171875 }, { "word": "选", "start": 297.0198557692308, "end": 297.2798557692308, "probability": 0.99951171875 }, { "word": "中", "start": 297.2798557692308, "end": 297.41985576923076, "probability": 0.998046875 }, { "word": "的", "start": 297.41985576923076, "end": 297.5198557692308, "probability": 1.0 }, { "word": "条", "start": 297.5198557692308, "end": 297.6598557692308, "probability": 1.0 }, { "word": "目", "start": 297.6598557692308, "end": 297.85985576923076, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.09651627260095932, "compression_ratio": 1.4588607594936709, "no_speech_prob": 2.7198640908943084e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和复用模式完整层会正常执行所有操作它会生成全局KV生成锁引起K运行吸收锁引起并选择自己的TopK记忆条目如果需要快速回顾一下吸收注意力吸收锁引起基本上就是KV缓存的一个低成本搜索引擎它会给缓存的全局KV条目打分选出最相关的TopK条目而注意力只读取这些被选中的条目接下来是重锁引模式该层会使用上一层已经生成的全局KV但它对哪些信息" } ], "id": 144, "review_annotation": "", "row_id": 144, "bounded_row_id": 144, "parent_row_ids": [ 144 ], "parent_row_id": 144, "packet_id": 10, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 275.0798557692308, "packet_end": 304.2678413461539, "raw_text": "而注意力只读取这些被选中的条目", "selected_text": "而注意力只读取这些被选中的条目", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 144 ], "packet_id": 10, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 297.85985576923076, "end": 300.09985576923077, "text": "接下来是重锁引模式", "raw_asr_text": "接下来是重锁引模式", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 10, "words": [ { "word": "接", "start": 297.85985576923076, "end": 298.1798557692308, "probability": 0.99853515625 }, { "word": "下", "start": 298.1798557692308, "end": 298.3398557692308, "probability": 0.99951171875 }, { "word": "来", "start": 298.3398557692308, "end": 298.53985576923077, "probability": 1.0 }, { "word": "是", "start": 298.53985576923077, "end": 298.8798557692308, "probability": 0.9951171875 }, { "word": "重", "start": 298.8798557692308, "end": 299.1998557692308, "probability": 0.75146484375 }, { "word": "锁", "start": 299.1998557692308, "end": 299.5198557692308, "probability": 0.998046875 }, { "word": "引", "start": 299.5198557692308, "end": 299.6798557692308, "probability": 0.9990234375 }, { "word": "模", "start": 299.6798557692308, "end": 299.85985576923076, "probability": 0.9912109375 }, { "word": "式", "start": 299.85985576923076, "end": 300.09985576923077, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09651627260095932, "compression_ratio": 1.4588607594936709, "no_speech_prob": 2.7198640908943084e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和复用模式完整层会正常执行所有操作它会生成全局KV生成锁引起K运行吸收锁引起并选择自己的TopK记忆条目如果需要快速回顾一下吸收注意力吸收锁引起基本上就是KV缓存的一个低成本搜索引擎它会给缓存的全局KV条目打分选出最相关的TopK条目而注意力只读取这些被选中的条目接下来是重锁引模式该层会使用上一层已经生成的全局KV但它对哪些信息" } ], "id": 145, "review_annotation": "", "row_id": 145, "bounded_row_id": 145, "parent_row_ids": [ 145 ], "parent_row_id": 145, "packet_id": 10, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 275.0798557692308, "packet_end": 304.2678413461539, "raw_text": "接下来是重锁引模式", "selected_text": "接下来是重锁引模式", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 145 ], "packet_id": 10, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 300.09985576923077, "end": 302.9998557692308, "text": "该层会使用上一层已经生成的全局KV", "raw_asr_text": "该层会使用上一层已经生成的全局KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 10, "words": [ { "word": "该", "start": 300.09985576923077, "end": 300.4598557692308, "probability": 0.9951171875 }, { "word": "层", "start": 300.4598557692308, "end": 300.6998557692308, "probability": 1.0 }, { "word": "会", "start": 300.6998557692308, "end": 300.8798557692308, "probability": 0.99951171875 }, { "word": "使", "start": 300.8798557692308, "end": 301.09985576923077, "probability": 0.99951171875 }, { "word": "用", "start": 301.09985576923077, "end": 301.1798557692308, "probability": 1.0 }, { "word": "上", "start": 301.1798557692308, "end": 301.3798557692308, "probability": 0.99951171875 }, { "word": "一", "start": 301.3798557692308, "end": 301.5198557692308, "probability": 0.998046875 }, { "word": "层", "start": 301.5198557692308, "end": 301.6598557692308, "probability": 1.0 }, { "word": "已经", "start": 301.6598557692308, "end": 301.8798557692308, "probability": 0.99755859375 }, { "word": "生", "start": 301.8798557692308, "end": 302.1398557692308, "probability": 0.99755859375 }, { "word": "成", "start": 302.1398557692308, "end": 302.2798557692308, "probability": 0.998046875 }, { "word": "的", "start": 302.2798557692308, "end": 302.3798557692308, "probability": 0.99951171875 }, { "word": "全", "start": 302.3798557692308, "end": 302.53985576923077, "probability": 0.99951171875 }, { "word": "局", "start": 302.53985576923077, "end": 302.6598557692308, "probability": 0.9990234375 }, { "word": "K", "start": 302.6598557692308, "end": 302.8198557692308, "probability": 0.99951171875 }, { "word": "V", "start": 302.8198557692308, "end": 302.9998557692308, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09651627260095932, "compression_ratio": 1.4588607594936709, "no_speech_prob": 2.7198640908943084e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和复用模式完整层会正常执行所有操作它会生成全局KV生成锁引起K运行吸收锁引起并选择自己的TopK记忆条目如果需要快速回顾一下吸收注意力吸收锁引起基本上就是KV缓存的一个低成本搜索引擎它会给缓存的全局KV条目打分选出最相关的TopK条目而注意力只读取这些被选中的条目接下来是重锁引模式该层会使用上一层已经生成的全局KV但它对哪些信息" } ], "id": 146, "review_annotation": "", "row_id": 146, "bounded_row_id": 146, "parent_row_ids": [ 146 ], "parent_row_id": 146, "packet_id": 10, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 275.0798557692308, "packet_end": 304.2678413461539, "raw_text": "该层会使用上一层已经生成的全局KV", "selected_text": "该层会使用上一层已经生成的全局KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 146 ], "packet_id": 10, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 303.4678413461539, "end": 306.04784134615386, "text": "但它对哪些信息重要可能有不同判断", "raw_asr_text": "但它对哪些信息重要可能有不同判断", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 11, "words": [ { "word": "但", "start": 302.66784134615386, "end": 303.2678413461539, "probability": 0.642578125 }, { "word": "它", "start": 303.2678413461539, "end": 303.42784134615385, "probability": 0.79150390625 }, { "word": "对", "start": 303.42784134615385, "end": 303.60784134615386, "probability": 0.98193359375 }, { "word": "哪", "start": 303.60784134615386, "end": 303.80784134615385, "probability": 0.96533203125 }, { "word": "些", "start": 303.80784134615385, "end": 303.98784134615386, "probability": 0.99560546875 }, { "word": "信", "start": 303.98784134615386, "end": 304.12784134615384, "probability": 0.9970703125 }, { "word": "息", "start": 304.12784134615384, "end": 304.3278413461539, "probability": 1.0 }, { "word": "重要", "start": 304.3278413461539, "end": 304.56784134615384, "probability": 0.9951171875 }, { "word": "可能", "start": 304.56784134615384, "end": 304.9678413461539, "probability": 0.97021484375 }, { "word": "有", "start": 304.9678413461539, "end": 305.22784134615387, "probability": 0.99462890625 }, { "word": "不同", "start": 305.22784134615387, "end": 305.54784134615386, "probability": 0.9990234375 }, { "word": "判", "start": 305.54784134615386, "end": 305.8278413461539, "probability": 0.96630859375 }, { "word": "断", "start": 305.8278413461539, "end": 306.04784134615386, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.10691480364118304, "compression_ratio": 1.206896551724138, "no_speech_prob": 3.035002110740592e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但它对哪些信息重要可能有不同判断因此它会附用上一层的主KV和索引器K但生成自己的query并重新选择TopK最后一种是附用模式这种模式会附用一切包括主KV和TopK索引甚至不需要再次运行索引器它只对上一层已经选中的内容执行注意力除了改造CSA2之外V4.1 Flash还移除了V4中那种在SWA与HCA或CSA之间所以" } ], "id": 147, "review_annotation": "", "row_id": 147, "bounded_row_id": 147, "parent_row_ids": [ 147 ], "parent_row_id": 147, "packet_id": 11, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 302.66784134615386, "packet_end": 331.85582692307696, "raw_text": "但它对哪些信息重要可能有不同判断", "selected_text": "但它对哪些信息重要可能有不同判断", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 147 ], "packet_id": 11, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 306.04784134615386, "end": 309.48784134615386, "text": "因此它会附用上一层的主KV和索引器K", "raw_asr_text": "因此它会附用上一层的主KV和索引器K", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 11, "words": [ { "word": "因", "start": 306.04784134615386, "end": 306.34784134615387, "probability": 0.96240234375 }, { "word": "此", "start": 306.34784134615387, "end": 306.5878413461539, "probability": 1.0 }, { "word": "它", "start": 306.5878413461539, "end": 306.7678413461539, "probability": 0.951171875 }, { "word": "会", "start": 306.7678413461539, "end": 306.92784134615385, "probability": 0.99658203125 }, { "word": "附", "start": 306.92784134615385, "end": 307.04784134615386, "probability": 0.9404296875 }, { "word": "用", "start": 307.04784134615386, "end": 307.18784134615385, "probability": 0.99853515625 }, { "word": "上", "start": 307.18784134615385, "end": 307.44784134615384, "probability": 0.99755859375 }, { "word": "一", "start": 307.44784134615384, "end": 307.66784134615386, "probability": 0.98828125 }, { "word": "层", "start": 307.66784134615386, "end": 307.84784134615387, "probability": 0.99951171875 }, { "word": "的", "start": 307.84784134615387, "end": 307.98784134615386, "probability": 0.98779296875 }, { "word": "主", "start": 307.98784134615386, "end": 308.18784134615385, "probability": 0.986328125 }, { "word": "K", "start": 308.18784134615385, "end": 308.3278413461539, "probability": 0.94287109375 }, { "word": "V", "start": 308.3278413461539, "end": 308.50784134615384, "probability": 0.978515625 }, { "word": "和", "start": 308.50784134615384, "end": 308.80784134615385, "probability": 0.92919921875 }, { "word": "索", "start": 308.80784134615385, "end": 309.0278413461539, "probability": 0.783203125 }, { "word": "引", "start": 309.0278413461539, "end": 309.12784134615384, "probability": 0.99658203125 }, { "word": "器", "start": 309.12784134615384, "end": 309.28784134615387, "probability": 0.9619140625 }, { "word": "K", "start": 309.28784134615387, "end": 309.48784134615386, "probability": 0.904296875 } ], "asr_metrics": { "avg_logprob": -0.10691480364118304, "compression_ratio": 1.206896551724138, "no_speech_prob": 3.035002110740592e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但它对哪些信息重要可能有不同判断因此它会附用上一层的主KV和索引器K但生成自己的query并重新选择TopK最后一种是附用模式这种模式会附用一切包括主KV和TopK索引甚至不需要再次运行索引器它只对上一层已经选中的内容执行注意力除了改造CSA2之外V4.1 Flash还移除了V4中那种在SWA与HCA或CSA之间所以" } ], "id": 148, "review_annotation": "", "row_id": 148, "bounded_row_id": 148, "parent_row_ids": [ 148 ], "parent_row_id": 148, "packet_id": 11, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 302.66784134615386, "packet_end": 331.85582692307696, "raw_text": "因此它会附用上一层的主KV和索引器K", "selected_text": "因此它会附用上一层的主KV和索引器K", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 148 ], "packet_id": 11, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 309.48784134615386, "end": 312.78784134615387, "text": "但生成自己的query并重新选择TopK", "raw_asr_text": "但生成自己的query并重新选择TopK", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 11, "words": [ { "word": "但", "start": 309.48784134615386, "end": 309.84784134615387, "probability": 0.91650390625 }, { "word": "生", "start": 309.84784134615387, "end": 310.04784134615386, "probability": 0.99072265625 }, { "word": "成", "start": 310.04784134615386, "end": 310.2078413461539, "probability": 0.994140625 }, { "word": "自己", "start": 310.2078413461539, "end": 310.54784134615386, "probability": 0.98681640625 }, { "word": "的", "start": 310.54784134615386, "end": 310.74784134615385, "probability": 0.998046875 }, { "word": "qu", "start": 310.74784134615385, "end": 310.92784134615385, "probability": 0.5126953125 }, { "word": "ery", "start": 310.92784134615385, "end": 311.12784134615384, "probability": 0.99267578125 }, { "word": "并", "start": 311.12784134615384, "end": 311.50784134615384, "probability": 0.9814453125 }, { "word": "重", "start": 311.50784134615384, "end": 311.7078413461539, "probability": 0.9873046875 }, { "word": "新", "start": 311.7078413461539, "end": 311.98784134615386, "probability": 0.99951171875 }, { "word": "选", "start": 311.98784134615386, "end": 312.2078413461539, "probability": 0.99951171875 }, { "word": "择", "start": 312.2078413461539, "end": 312.3878413461539, "probability": 0.9990234375 }, { "word": "Top", "start": 312.3878413461539, "end": 312.5878413461539, "probability": 0.64111328125 }, { "word": "K", "start": 312.5878413461539, "end": 312.78784134615387, "probability": 0.83544921875 } ], "asr_metrics": { "avg_logprob": -0.10691480364118304, "compression_ratio": 1.206896551724138, "no_speech_prob": 3.035002110740592e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但它对哪些信息重要可能有不同判断因此它会附用上一层的主KV和索引器K但生成自己的query并重新选择TopK最后一种是附用模式这种模式会附用一切包括主KV和TopK索引甚至不需要再次运行索引器它只对上一层已经选中的内容执行注意力除了改造CSA2之外V4.1 Flash还移除了V4中那种在SWA与HCA或CSA之间所以" } ], "id": 149, "review_annotation": "", "row_id": 149, "bounded_row_id": 149, "parent_row_ids": [ 149 ], "parent_row_id": 149, "packet_id": 11, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 302.66784134615386, "packet_end": 331.85582692307696, "raw_text": "但生成自己的query并重新选择TopK", "selected_text": "但生成自己的query并重新选择TopK", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 149 ], "packet_id": 11, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 312.78784134615387, "end": 314.36784134615385, "text": "最后一种是附用模式", "raw_asr_text": "最后一种是附用模式", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 11, "words": [ { "word": "最", "start": 312.78784134615387, "end": 313.1478413461539, "probability": 0.9501953125 }, { "word": "后", "start": 313.1478413461539, "end": 313.28784134615387, "probability": 0.99951171875 }, { "word": "一", "start": 313.28784134615387, "end": 313.3878413461539, "probability": 0.9951171875 }, { "word": "种", "start": 313.3878413461539, "end": 313.5278413461539, "probability": 0.9990234375 }, { "word": "是", "start": 313.5278413461539, "end": 313.74784134615385, "probability": 0.9921875 }, { "word": "附", "start": 313.74784134615385, "end": 313.94784134615384, "probability": 0.9912109375 }, { "word": "用", "start": 313.94784134615384, "end": 314.04784134615386, "probability": 0.99853515625 }, { "word": "模", "start": 314.04784134615386, "end": 314.18784134615385, "probability": 0.99755859375 }, { "word": "式", "start": 314.18784134615385, "end": 314.36784134615385, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10691480364118304, "compression_ratio": 1.206896551724138, "no_speech_prob": 3.035002110740592e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但它对哪些信息重要可能有不同判断因此它会附用上一层的主KV和索引器K但生成自己的query并重新选择TopK最后一种是附用模式这种模式会附用一切包括主KV和TopK索引甚至不需要再次运行索引器它只对上一层已经选中的内容执行注意力除了改造CSA2之外V4.1 Flash还移除了V4中那种在SWA与HCA或CSA之间所以" } ], "id": 150, "review_annotation": "", "row_id": 150, "bounded_row_id": 150, "parent_row_ids": [ 150 ], "parent_row_id": 150, "packet_id": 11, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 302.66784134615386, "packet_end": 331.85582692307696, "raw_text": "最后一种是附用模式", "selected_text": "最后一种是附用模式", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 150 ], "packet_id": 11, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 314.36784134615385, "end": 316.36784134615385, "text": "这种模式会附用一切", "raw_asr_text": "这种模式会附用一切", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 11, "words": [ { "word": "这", "start": 314.36784134615385, "end": 314.84784134615387, "probability": 0.9853515625 }, { "word": "种", "start": 314.84784134615387, "end": 315.00784134615384, "probability": 0.99853515625 }, { "word": "模", "start": 315.00784134615384, "end": 315.18784134615385, "probability": 0.99658203125 }, { "word": "式", "start": 315.18784134615385, "end": 315.3878413461539, "probability": 1.0 }, { "word": "会", "start": 315.3878413461539, "end": 315.62784134615384, "probability": 0.998046875 }, { "word": "附", "start": 315.62784134615384, "end": 315.7678413461539, "probability": 0.9990234375 }, { "word": "用", "start": 315.7678413461539, "end": 315.9078413461539, "probability": 0.9990234375 }, { "word": "一", "start": 315.9078413461539, "end": 316.10784134615386, "probability": 0.99609375 }, { "word": "切", "start": 316.10784134615386, "end": 316.36784134615385, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.10691480364118304, "compression_ratio": 1.206896551724138, "no_speech_prob": 3.035002110740592e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但它对哪些信息重要可能有不同判断因此它会附用上一层的主KV和索引器K但生成自己的query并重新选择TopK最后一种是附用模式这种模式会附用一切包括主KV和TopK索引甚至不需要再次运行索引器它只对上一层已经选中的内容执行注意力除了改造CSA2之外V4.1 Flash还移除了V4中那种在SWA与HCA或CSA之间所以" } ], "id": 151, "review_annotation": "", "row_id": 151, "bounded_row_id": 151, "parent_row_ids": [ 151 ], "parent_row_id": 151, "packet_id": 11, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 302.66784134615386, "packet_end": 331.85582692307696, "raw_text": "这种模式会附用一切", "selected_text": "这种模式会附用一切", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 151 ], "packet_id": 11, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 316.36784134615385, "end": 318.5278413461539, "text": "包括主KV和TopK索引", "raw_asr_text": "包括主KV和TopK索引", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 11, "words": [ { "word": "包括", "start": 316.36784134615385, "end": 316.7678413461539, "probability": 0.97021484375 }, { "word": "主", "start": 316.7678413461539, "end": 316.98784134615386, "probability": 0.99169921875 }, { "word": "K", "start": 316.98784134615386, "end": 317.16784134615386, "probability": 0.9970703125 }, { "word": "V", "start": 317.16784134615386, "end": 317.3878413461539, "probability": 0.998046875 }, { "word": "和", "start": 317.3878413461539, "end": 317.6478413461539, "probability": 0.966796875 }, { "word": "Top", "start": 317.6478413461539, "end": 317.84784134615387, "probability": 0.994140625 }, { "word": "K", "start": 317.84784134615387, "end": 318.06784134615384, "probability": 0.99658203125 }, { "word": "索", "start": 318.06784134615384, "end": 318.3278413461539, "probability": 0.98095703125 }, { "word": "引", "start": 318.3278413461539, "end": 318.5278413461539, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10691480364118304, "compression_ratio": 1.206896551724138, "no_speech_prob": 3.035002110740592e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但它对哪些信息重要可能有不同判断因此它会附用上一层的主KV和索引器K但生成自己的query并重新选择TopK最后一种是附用模式这种模式会附用一切包括主KV和TopK索引甚至不需要再次运行索引器它只对上一层已经选中的内容执行注意力除了改造CSA2之外V4.1 Flash还移除了V4中那种在SWA与HCA或CSA之间所以" } ], "id": 152, "review_annotation": "", "row_id": 152, "bounded_row_id": 152, "parent_row_ids": [ 152 ], "parent_row_id": 152, "packet_id": 11, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 302.66784134615386, "packet_end": 331.85582692307696, "raw_text": "包括主KV和TopK索引", "selected_text": "包括主KV和TopK索引", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 152 ], "packet_id": 11, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 318.5278413461539, "end": 321.2078413461539, "text": "甚至不需要再次运行索引器", "raw_asr_text": "甚至不需要再次运行索引器", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 11, "words": [ { "word": "甚", "start": 318.5278413461539, "end": 318.9078413461539, "probability": 0.98828125 }, { "word": "至", "start": 318.9078413461539, "end": 319.1478413461539, "probability": 1.0 }, { "word": "不", "start": 319.1478413461539, "end": 319.3878413461539, "probability": 0.98779296875 }, { "word": "需要", "start": 319.3878413461539, "end": 319.62784134615384, "probability": 0.99560546875 }, { "word": "再", "start": 319.62784134615384, "end": 319.94784134615384, "probability": 0.99658203125 }, { "word": "次", "start": 319.94784134615384, "end": 320.16784134615386, "probability": 0.99658203125 }, { "word": "运", "start": 320.16784134615386, "end": 320.36784134615385, "probability": 0.9990234375 }, { "word": "行", "start": 320.36784134615385, "end": 320.5278413461539, "probability": 0.9970703125 }, { "word": "索", "start": 320.5278413461539, "end": 320.80784134615385, "probability": 0.9287109375 }, { "word": "引", "start": 320.80784134615385, "end": 321.00784134615384, "probability": 1.0 }, { "word": "器", "start": 321.00784134615384, "end": 321.2078413461539, "probability": 0.99462890625 } ], "asr_metrics": { "avg_logprob": -0.10691480364118304, "compression_ratio": 1.206896551724138, "no_speech_prob": 3.035002110740592e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但它对哪些信息重要可能有不同判断因此它会附用上一层的主KV和索引器K但生成自己的query并重新选择TopK最后一种是附用模式这种模式会附用一切包括主KV和TopK索引甚至不需要再次运行索引器它只对上一层已经选中的内容执行注意力除了改造CSA2之外V4.1 Flash还移除了V4中那种在SWA与HCA或CSA之间所以" } ], "id": 153, "review_annotation": "", "row_id": 153, "bounded_row_id": 153, "parent_row_ids": [ 153 ], "parent_row_id": 153, "packet_id": 11, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 302.66784134615386, "packet_end": 331.85582692307696, "raw_text": "甚至不需要再次运行索引器", "selected_text": "甚至不需要再次运行索引器", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 153 ], "packet_id": 11, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 321.2078413461539, "end": 324.42784134615385, "text": "它只对上一层已经选中的内容执行注意力", "raw_asr_text": "它只对上一层已经选中的内容执行注意力", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 11, "words": [ { "word": "它", "start": 321.2078413461539, "end": 321.68784134615385, "probability": 0.97119140625 }, { "word": "只", "start": 321.68784134615385, "end": 321.8278413461539, "probability": 0.9970703125 }, { "word": "对", "start": 321.8278413461539, "end": 321.9678413461539, "probability": 0.99365234375 }, { "word": "上", "start": 321.9678413461539, "end": 322.1478413461539, "probability": 0.998046875 }, { "word": "一", "start": 322.1478413461539, "end": 322.28784134615387, "probability": 0.986328125 }, { "word": "层", "start": 322.28784134615387, "end": 322.50784134615384, "probability": 1.0 }, { "word": "已经", "start": 322.50784134615384, "end": 322.8278413461539, "probability": 0.99169921875 }, { "word": "选", "start": 322.8278413461539, "end": 323.18784134615385, "probability": 1.0 }, { "word": "中", "start": 323.18784134615385, "end": 323.3278413461539, "probability": 0.99853515625 }, { "word": "的", "start": 323.3278413461539, "end": 323.44784134615384, "probability": 0.9970703125 }, { "word": "内", "start": 323.44784134615384, "end": 323.5878413461539, "probability": 0.9970703125 }, { "word": "容", "start": 323.5878413461539, "end": 323.72784134615387, "probability": 0.99951171875 }, { "word": "执", "start": 323.72784134615387, "end": 323.9078413461539, "probability": 0.9970703125 }, { "word": "行", "start": 323.9078413461539, "end": 324.0278413461539, "probability": 0.99951171875 }, { "word": "注意", "start": 324.0278413461539, "end": 324.22784134615387, "probability": 0.99560546875 }, { "word": "力", "start": 324.22784134615387, "end": 324.42784134615385, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.10691480364118304, "compression_ratio": 1.206896551724138, "no_speech_prob": 3.035002110740592e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但它对哪些信息重要可能有不同判断因此它会附用上一层的主KV和索引器K但生成自己的query并重新选择TopK最后一种是附用模式这种模式会附用一切包括主KV和TopK索引甚至不需要再次运行索引器它只对上一层已经选中的内容执行注意力除了改造CSA2之外V4.1 Flash还移除了V4中那种在SWA与HCA或CSA之间所以" } ], "id": 154, "review_annotation": "", "row_id": 154, "bounded_row_id": 154, "parent_row_ids": [ 154 ], "parent_row_id": 154, "packet_id": 11, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 302.66784134615386, "packet_end": 331.85582692307696, "raw_text": "它只对上一层已经选中的内容执行注意力", "selected_text": "它只对上一层已经选中的内容执行注意力", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 154 ], "packet_id": 11, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 324.42784134615385, "end": 326.36784134615385, "text": "除了改造CSA2之外", "raw_asr_text": "除了改造CSA2之外", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 11, "words": [ { "word": "除", "start": 324.42784134615385, "end": 324.9678413461539, "probability": 0.9951171875 }, { "word": "了", "start": 324.9678413461539, "end": 325.10784134615386, "probability": 0.9990234375 }, { "word": "改", "start": 325.10784134615386, "end": 325.28784134615387, "probability": 0.99755859375 }, { "word": "造", "start": 325.28784134615387, "end": 325.48784134615386, "probability": 0.99853515625 }, { "word": "C", "start": 325.48784134615386, "end": 325.66784134615386, "probability": 0.94921875 }, { "word": "SA", "start": 325.66784134615386, "end": 325.92784134615385, "probability": 0.9833984375 }, { "word": "2", "start": 325.92784134615385, "end": 326.0878413461539, "probability": 0.63427734375 }, { "word": "之", "start": 326.0878413461539, "end": 326.2078413461539, "probability": 0.984375 }, { "word": "外", "start": 326.2078413461539, "end": 326.36784134615385, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.10691480364118304, "compression_ratio": 1.206896551724138, "no_speech_prob": 3.035002110740592e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但它对哪些信息重要可能有不同判断因此它会附用上一层的主KV和索引器K但生成自己的query并重新选择TopK最后一种是附用模式这种模式会附用一切包括主KV和TopK索引甚至不需要再次运行索引器它只对上一层已经选中的内容执行注意力除了改造CSA2之外V4.1 Flash还移除了V4中那种在SWA与HCA或CSA之间所以" } ], "id": 155, "review_annotation": "", "row_id": 155, "bounded_row_id": 155, "parent_row_ids": [ 155 ], "parent_row_id": 155, "packet_id": 11, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 302.66784134615386, "packet_end": 331.85582692307696, "raw_text": "除了改造CSA2之外", "selected_text": "除了改造CSA2之外", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 155 ], "packet_id": 11, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 326.36784134615385, "end": 328.94784134615384, "text": "V4.1 Flash还移除了V4中那种", "raw_asr_text": "V4.1 Flash还移除了V4中那种", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 11, "words": [ { "word": "V", "start": 326.36784134615385, "end": 326.60784134615386, "probability": 0.736328125 }, { "word": "4.", "start": 326.60784134615386, "end": 326.7678413461539, "probability": 0.9970703125 }, { "word": "1", "start": 326.8878413461539, "end": 327.06784134615384, "probability": 0.91552734375 }, { "word": " Flash", "start": 327.06784134615384, "end": 327.36784134615385, "probability": 0.37109375 }, { "word": "还", "start": 327.36784134615385, "end": 327.68784134615385, "probability": 0.97802734375 }, { "word": "移", "start": 327.68784134615385, "end": 327.8878413461539, "probability": 0.99755859375 }, { "word": "除", "start": 327.8878413461539, "end": 328.0278413461539, "probability": 0.99951171875 }, { "word": "了", "start": 328.0278413461539, "end": 328.1478413461539, "probability": 0.99609375 }, { "word": "V", "start": 328.1478413461539, "end": 328.28784134615387, "probability": 0.98046875 }, { "word": "4", "start": 328.28784134615387, "end": 328.4678413461539, "probability": 0.9970703125 }, { "word": "中", "start": 328.4678413461539, "end": 328.60784134615386, "probability": 0.95263671875 }, { "word": "那", "start": 328.60784134615386, "end": 328.78784134615387, "probability": 0.9208984375 }, { "word": "种", "start": 328.78784134615387, "end": 328.94784134615384, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.10691480364118304, "compression_ratio": 1.206896551724138, "no_speech_prob": 3.035002110740592e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但它对哪些信息重要可能有不同判断因此它会附用上一层的主KV和索引器K但生成自己的query并重新选择TopK最后一种是附用模式这种模式会附用一切包括主KV和TopK索引甚至不需要再次运行索引器它只对上一层已经选中的内容执行注意力除了改造CSA2之外V4.1 Flash还移除了V4中那种在SWA与HCA或CSA之间所以" } ], "id": 156, "review_annotation": "", "row_id": 156, "bounded_row_id": 156, "parent_row_ids": [ 156 ], "parent_row_id": 156, "packet_id": 11, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 302.66784134615386, "packet_end": 331.85582692307696, "raw_text": "V4.1 Flash还移除了V4中那种", "selected_text": "V4.1 Flash还移除了V4中那种", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 156 ], "packet_id": 11, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 328.94784134615384, "end": 331.05582692307695, "text": "在SWA与HCA或CSA之间", "raw_asr_text": "在SWA与HCA或CSA之间", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 11, "words": [ { "word": "在", "start": 328.94784134615384, "end": 329.22784134615387, "probability": 0.98046875 }, { "word": "S", "start": 329.22784134615387, "end": 329.5278413461539, "probability": 0.970703125 }, { "word": "WA", "start": 329.5278413461539, "end": 329.98784134615386, "probability": 0.99072265625 }, { "word": "与", "start": 329.98784134615386, "end": 330.1478413461539, "probability": 0.943359375 }, { "word": "H", "start": 330.1478413461539, "end": 330.4078413461539, "probability": 0.9970703125 }, { "word": "CA", "start": 330.4078413461539, "end": 330.72784134615387, "probability": 0.9892578125 }, { "word": "或", "start": 330.72784134615387, "end": 330.8878413461539, "probability": 0.90478515625 }, { "word": "C", "start": 330.8878413461539, "end": 331.0878413461539, "probability": 0.9970703125 }, { "word": "SA", "start": 331.0878413461539, "end": 331.4078413461539, "probability": 0.99951171875 }, { "word": "之", "start": 331.4078413461539, "end": 331.5878413461539, "probability": 0.9921875 }, { "word": "间", "start": 331.5878413461539, "end": 331.72784134615387, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.10691480364118304, "compression_ratio": 1.206896551724138, "no_speech_prob": 3.035002110740592e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "但它对哪些信息重要可能有不同判断因此它会附用上一层的主KV和索引器K但生成自己的query并重新选择TopK最后一种是附用模式这种模式会附用一切包括主KV和TopK索引甚至不需要再次运行索引器它只对上一层已经选中的内容执行注意力除了改造CSA2之外V4.1 Flash还移除了V4中那种在SWA与HCA或CSA之间所以" } ], "id": 157, "review_annotation": "", "row_id": 157, "bounded_row_id": 157, "parent_row_ids": [ 157 ], "parent_row_id": 157, "packet_id": 11, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 302.66784134615386, "packet_end": 331.85582692307696, "raw_text": "在SWA与HCA或CSA之间", "selected_text": "在SWA与HCA或CSA之间", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 157 ], "packet_id": 11, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 331.05582692307695, "end": 333.55582692307695, "text": "HCA或CSA之间交错的混乱混合注意力", "raw_asr_text": "HCA或CSA之间交错的混乱混合注意力", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 12, "words": [ { "word": "H", "start": 330.25582692307694, "end": 330.41582692307696, "probability": 0.908203125 }, { "word": "CA", "start": 330.41582692307696, "end": 330.7158269230769, "probability": 0.99267578125 }, { "word": "或", "start": 330.7158269230769, "end": 330.8958269230769, "probability": 0.7724609375 }, { "word": "C", "start": 330.8958269230769, "end": 331.0958269230769, "probability": 0.9560546875 }, { "word": "SA", "start": 331.0958269230769, "end": 331.37582692307694, "probability": 0.99462890625 }, { "word": "之", "start": 331.37582692307694, "end": 331.5958269230769, "probability": 0.99267578125 }, { "word": "间", "start": 331.5958269230769, "end": 331.73582692307696, "probability": 0.9423828125 }, { "word": "交", "start": 331.73582692307696, "end": 331.97582692307697, "probability": 0.9990234375 }, { "word": "错", "start": 331.97582692307697, "end": 332.1558269230769, "probability": 0.9951171875 }, { "word": "的", "start": 332.1558269230769, "end": 332.2758269230769, "probability": 0.9921875 }, { "word": "混", "start": 332.2758269230769, "end": 332.49582692307695, "probability": 0.982421875 }, { "word": "乱", "start": 332.49582692307695, "end": 332.75582692307694, "probability": 0.99853515625 }, { "word": "混", "start": 332.75582692307694, "end": 332.9558269230769, "probability": 0.9853515625 }, { "word": "合", "start": 332.9558269230769, "end": 333.19582692307694, "probability": 0.9873046875 }, { "word": "注意", "start": 333.19582692307694, "end": 333.37582692307694, "probability": 0.9921875 }, { "word": "力", "start": 333.37582692307694, "end": 333.55582692307695, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09965962779765226, "compression_ratio": 1.1941489361702127, "no_speech_prob": 2.60604888363547e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "HCA或CSA之间交错的混乱混合注意力它们的主要改动是彻底弃用HCA于是全聚分支现在只剩纯CSAR从而得到更简单的CSAR SWA混合架构但即便CSAR已经减少了跨层所需存储的KV量仍存在另一个相当巨大的低效问题因为无论预填充阶段的注意力被压缩到什么程度模型人必须先让所有这些输入Token通过网络才能生成第一个Token设想你给一个常规40层Transformer" } ], "id": 158, "review_annotation": "", "row_id": 158, "bounded_row_id": 158, "parent_row_ids": [ 158 ], "parent_row_id": 158, "packet_id": 12, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 330.25582692307694, "packet_end": 359.44381250000004, "raw_text": "HCA或CSA之间交错的混乱混合注意力", "selected_text": "HCA或CSA之间交错的混乱混合注意力", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 158 ], "packet_id": 12, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 333.55582692307695, "end": 335.99582692307695, "text": "它们的主要改动是彻底弃用HCA", "raw_asr_text": "它们的主要改动是彻底弃用HCA", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 12, "words": [ { "word": "它", "start": 333.55582692307695, "end": 333.7758269230769, "probability": 0.64697265625 }, { "word": "们", "start": 333.7758269230769, "end": 333.93582692307695, "probability": 0.9990234375 }, { "word": "的", "start": 333.93582692307695, "end": 334.0358269230769, "probability": 0.99072265625 }, { "word": "主", "start": 334.0358269230769, "end": 334.1558269230769, "probability": 0.99951171875 }, { "word": "要", "start": 334.1558269230769, "end": 334.25582692307694, "probability": 1.0 }, { "word": "改", "start": 334.25582692307694, "end": 334.43582692307695, "probability": 0.99365234375 }, { "word": "动", "start": 334.43582692307695, "end": 334.5958269230769, "probability": 0.9951171875 }, { "word": "是", "start": 334.5958269230769, "end": 334.79582692307696, "probability": 0.99169921875 }, { "word": "彻", "start": 334.79582692307696, "end": 334.97582692307697, "probability": 0.99609375 }, { "word": "底", "start": 334.97582692307697, "end": 335.11582692307695, "probability": 1.0 }, { "word": "弃", "start": 335.11582692307695, "end": 335.2758269230769, "probability": 0.7060546875 }, { "word": "用", "start": 335.2758269230769, "end": 335.4558269230769, "probability": 0.998046875 }, { "word": "H", "start": 335.4558269230769, "end": 335.67582692307695, "probability": 0.984375 }, { "word": "CA", "start": 335.67582692307695, "end": 335.99582692307695, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.09965962779765226, "compression_ratio": 1.1941489361702127, "no_speech_prob": 2.60604888363547e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "HCA或CSA之间交错的混乱混合注意力它们的主要改动是彻底弃用HCA于是全聚分支现在只剩纯CSAR从而得到更简单的CSAR SWA混合架构但即便CSAR已经减少了跨层所需存储的KV量仍存在另一个相当巨大的低效问题因为无论预填充阶段的注意力被压缩到什么程度模型人必须先让所有这些输入Token通过网络才能生成第一个Token设想你给一个常规40层Transformer" } ], "id": 159, "review_annotation": "", "row_id": 159, "bounded_row_id": 159, "parent_row_ids": [ 159 ], "parent_row_id": 159, "packet_id": 12, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 330.25582692307694, "packet_end": 359.44381250000004, "raw_text": "它们的主要改动是彻底弃用HCA", "selected_text": "它们的主要改动是彻底弃用HCA", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 159 ], "packet_id": 12, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 335.99582692307695, "end": 338.99582692307695, "text": "于是全聚分支现在只剩纯CSAR", "raw_asr_text": "于是全聚分支现在只剩纯CSAR", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 12, "words": [ { "word": "于", "start": 335.99582692307695, "end": 336.17582692307695, "probability": 0.9775390625 }, { "word": "是", "start": 336.17582692307695, "end": 336.35582692307696, "probability": 0.99951171875 }, { "word": "全", "start": 336.35582692307696, "end": 336.63582692307693, "probability": 0.98828125 }, { "word": "聚", "start": 336.63582692307693, "end": 336.7758269230769, "probability": 0.8271484375 }, { "word": "分", "start": 336.7758269230769, "end": 336.9558269230769, "probability": 0.99853515625 }, { "word": "支", "start": 336.9558269230769, "end": 337.11582692307695, "probability": 0.9931640625 }, { "word": "现在", "start": 337.11582692307695, "end": 337.47582692307697, "probability": 0.95947265625 }, { "word": "只", "start": 337.47582692307697, "end": 337.85582692307696, "probability": 0.99853515625 }, { "word": "剩", "start": 337.85582692307696, "end": 338.05582692307695, "probability": 0.99853515625 }, { "word": "纯", "start": 338.05582692307695, "end": 338.31582692307694, "probability": 0.9892578125 }, { "word": "CS", "start": 338.31582692307694, "end": 338.7158269230769, "probability": 0.69287109375 }, { "word": "AR", "start": 338.7158269230769, "end": 338.99582692307695, "probability": 0.9970703125 } ], "asr_metrics": { "avg_logprob": -0.09965962779765226, "compression_ratio": 1.1941489361702127, "no_speech_prob": 2.60604888363547e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "HCA或CSA之间交错的混乱混合注意力它们的主要改动是彻底弃用HCA于是全聚分支现在只剩纯CSAR从而得到更简单的CSAR SWA混合架构但即便CSAR已经减少了跨层所需存储的KV量仍存在另一个相当巨大的低效问题因为无论预填充阶段的注意力被压缩到什么程度模型人必须先让所有这些输入Token通过网络才能生成第一个Token设想你给一个常规40层Transformer" } ], "id": 160, "review_annotation": "", "row_id": 160, "bounded_row_id": 160, "parent_row_ids": [ 160 ], "parent_row_id": 160, "packet_id": 12, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 330.25582692307694, "packet_end": 359.44381250000004, "raw_text": "于是全聚分支现在只剩纯CSAR", "selected_text": "于是全聚分支现在只剩纯CSAR", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 160 ], "packet_id": 12, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 338.99582692307695, "end": 342.69582692307694, "text": "从而得到更简单的CSAR SWA混合架构", "raw_asr_text": "从而得到更简单的CSAR SWA混合架构", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 12, "words": [ { "word": "从", "start": 338.99582692307695, "end": 339.35582692307696, "probability": 0.9765625 }, { "word": "而", "start": 339.35582692307696, "end": 339.4558269230769, "probability": 1.0 }, { "word": "得", "start": 339.4558269230769, "end": 339.55582692307695, "probability": 0.9990234375 }, { "word": "到", "start": 339.55582692307695, "end": 339.75582692307694, "probability": 0.9990234375 }, { "word": "更", "start": 339.75582692307694, "end": 339.93582692307695, "probability": 0.99755859375 }, { "word": "简", "start": 339.93582692307695, "end": 340.19582692307694, "probability": 1.0 }, { "word": "单", "start": 340.19582692307694, "end": 340.35582692307696, "probability": 1.0 }, { "word": "的", "start": 340.35582692307696, "end": 340.49582692307695, "probability": 0.99755859375 }, { "word": "CS", "start": 340.49582692307695, "end": 340.8958269230769, "probability": 0.95263671875 }, { "word": "AR", "start": 340.8958269230769, "end": 341.2158269230769, "probability": 0.947265625 }, { "word": " S", "start": 341.2158269230769, "end": 341.4558269230769, "probability": 0.56103515625 }, { "word": "WA", "start": 341.4558269230769, "end": 341.8958269230769, "probability": 0.990234375 }, { "word": "混", "start": 341.8958269230769, "end": 342.13582692307693, "probability": 0.99560546875 }, { "word": "合", "start": 342.13582692307693, "end": 342.35582692307696, "probability": 0.99853515625 }, { "word": "架", "start": 342.35582692307696, "end": 342.51582692307693, "probability": 0.998046875 }, { "word": "构", "start": 342.51582692307693, "end": 342.69582692307694, "probability": 0.9970703125 } ], "asr_metrics": { "avg_logprob": -0.09965962779765226, "compression_ratio": 1.1941489361702127, "no_speech_prob": 2.60604888363547e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "HCA或CSA之间交错的混乱混合注意力它们的主要改动是彻底弃用HCA于是全聚分支现在只剩纯CSAR从而得到更简单的CSAR SWA混合架构但即便CSAR已经减少了跨层所需存储的KV量仍存在另一个相当巨大的低效问题因为无论预填充阶段的注意力被压缩到什么程度模型人必须先让所有这些输入Token通过网络才能生成第一个Token设想你给一个常规40层Transformer" } ], "id": 161, "review_annotation": "", "row_id": 161, "bounded_row_id": 161, "parent_row_ids": [ 161 ], "parent_row_id": 161, "packet_id": 12, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 330.25582692307694, "packet_end": 359.44381250000004, "raw_text": "从而得到更简单的CSAR SWA混合架构", "selected_text": "从而得到更简单的CSAR SWA混合架构", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 161 ], "packet_id": 12, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 342.69582692307694, "end": 346.8358269230769, "text": "但即便CSAR已经减少了跨层所需存储的KV量", "raw_asr_text": "但即便CSAR已经减少了跨层所需存储的KV量", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 12, "words": [ { "word": "但", "start": 342.69582692307694, "end": 342.97582692307697, "probability": 0.99609375 }, { "word": "即", "start": 342.97582692307697, "end": 343.19582692307694, "probability": 0.99951171875 }, { "word": "便", "start": 343.19582692307694, "end": 343.41582692307696, "probability": 0.99951171875 }, { "word": "CS", "start": 343.41582692307696, "end": 343.73582692307696, "probability": 0.984375 }, { "word": "AR", "start": 343.73582692307696, "end": 343.99582692307695, "probability": 0.9990234375 }, { "word": "已经", "start": 343.99582692307695, "end": 344.29582692307696, "probability": 0.97705078125 }, { "word": "减", "start": 344.29582692307696, "end": 344.5358269230769, "probability": 0.9990234375 }, { "word": "少", "start": 344.5358269230769, "end": 344.69582692307694, "probability": 0.99951171875 }, { "word": "了", "start": 344.69582692307694, "end": 344.8358269230769, "probability": 0.99755859375 }, { "word": "跨", "start": 344.8358269230769, "end": 345.11582692307695, "probability": 0.99462890625 }, { "word": "层", "start": 345.11582692307695, "end": 345.35582692307696, "probability": 0.9970703125 }, { "word": "所", "start": 345.35582692307696, "end": 345.63582692307693, "probability": 0.99462890625 }, { "word": "需", "start": 345.63582692307693, "end": 345.8358269230769, "probability": 0.99658203125 }, { "word": "存", "start": 345.8358269230769, "end": 346.05582692307695, "probability": 0.998046875 }, { "word": "储", "start": 346.05582692307695, "end": 346.17582692307695, "probability": 0.9990234375 }, { "word": "的", "start": 346.17582692307695, "end": 346.29582692307696, "probability": 0.99853515625 }, { "word": "K", "start": 346.29582692307696, "end": 346.4558269230769, "probability": 0.7900390625 }, { "word": "V", "start": 346.4558269230769, "end": 346.61582692307695, "probability": 0.99609375 }, { "word": "量", "start": 346.61582692307695, "end": 346.8358269230769, "probability": 0.970703125 } ], "asr_metrics": { "avg_logprob": -0.09965962779765226, "compression_ratio": 1.1941489361702127, "no_speech_prob": 2.60604888363547e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "HCA或CSA之间交错的混乱混合注意力它们的主要改动是彻底弃用HCA于是全聚分支现在只剩纯CSAR从而得到更简单的CSAR SWA混合架构但即便CSAR已经减少了跨层所需存储的KV量仍存在另一个相当巨大的低效问题因为无论预填充阶段的注意力被压缩到什么程度模型人必须先让所有这些输入Token通过网络才能生成第一个Token设想你给一个常规40层Transformer" } ], "id": 162, "review_annotation": "", "row_id": 162, "bounded_row_id": 162, "parent_row_ids": [ 162 ], "parent_row_id": 162, "packet_id": 12, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 330.25582692307694, "packet_end": 359.44381250000004, "raw_text": "但即便CSAR已经减少了跨层所需存储的KV量", "selected_text": "但即便CSAR已经减少了跨层所需存储的KV量", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 162 ], "packet_id": 12, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 346.8358269230769, "end": 349.47582692307697, "text": "仍存在另一个相当巨大的低效问题", "raw_asr_text": "仍存在另一个相当巨大的低效问题", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 12, "words": [ { "word": "仍", "start": 346.8358269230769, "end": 347.2758269230769, "probability": 0.99755859375 }, { "word": "存", "start": 347.2758269230769, "end": 347.51582692307693, "probability": 0.99853515625 }, { "word": "在", "start": 347.51582692307693, "end": 347.69582692307694, "probability": 0.99951171875 }, { "word": "另", "start": 347.69582692307694, "end": 347.85582692307696, "probability": 0.9892578125 }, { "word": "一个", "start": 347.85582692307696, "end": 348.07582692307693, "probability": 0.99609375 }, { "word": "相", "start": 348.07582692307693, "end": 348.2758269230769, "probability": 0.9970703125 }, { "word": "当", "start": 348.2758269230769, "end": 348.43582692307695, "probability": 0.9990234375 }, { "word": "巨", "start": 348.43582692307695, "end": 348.63582692307693, "probability": 0.998046875 }, { "word": "大的", "start": 348.63582692307693, "end": 348.8358269230769, "probability": 0.99560546875 }, { "word": "低", "start": 348.8358269230769, "end": 348.99582692307695, "probability": 0.97216796875 }, { "word": "效", "start": 348.99582692307695, "end": 349.13582692307693, "probability": 0.9990234375 }, { "word": "问题", "start": 349.13582692307693, "end": 349.47582692307697, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.09965962779765226, "compression_ratio": 1.1941489361702127, "no_speech_prob": 2.60604888363547e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "HCA或CSA之间交错的混乱混合注意力它们的主要改动是彻底弃用HCA于是全聚分支现在只剩纯CSAR从而得到更简单的CSAR SWA混合架构但即便CSAR已经减少了跨层所需存储的KV量仍存在另一个相当巨大的低效问题因为无论预填充阶段的注意力被压缩到什么程度模型人必须先让所有这些输入Token通过网络才能生成第一个Token设想你给一个常规40层Transformer" } ], "id": 163, "review_annotation": "", "row_id": 163, "bounded_row_id": 163, "parent_row_ids": [ 163 ], "parent_row_id": 163, "packet_id": 12, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 330.25582692307694, "packet_end": 359.44381250000004, "raw_text": "仍存在另一个相当巨大的低效问题", "selected_text": "仍存在另一个相当巨大的低效问题", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 163 ], "packet_id": 12, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 349.47582692307697, "end": 352.91582692307696, "text": "因为无论预填充阶段的注意力被压缩到什么程度", "raw_asr_text": "因为无论预填充阶段的注意力被压缩到什么程度", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 12, "words": [ { "word": "因为", "start": 349.47582692307697, "end": 350.07582692307693, "probability": 0.99072265625 }, { "word": "无", "start": 350.07582692307693, "end": 350.3358269230769, "probability": 0.9970703125 }, { "word": "论", "start": 350.3358269230769, "end": 350.49582692307695, "probability": 1.0 }, { "word": "预", "start": 350.49582692307695, "end": 350.6558269230769, "probability": 0.9365234375 }, { "word": "填", "start": 350.6558269230769, "end": 350.7758269230769, "probability": 0.78076171875 }, { "word": "充", "start": 350.7758269230769, "end": 350.8958269230769, "probability": 0.99609375 }, { "word": "阶", "start": 350.8958269230769, "end": 351.0358269230769, "probability": 0.9970703125 }, { "word": "段", "start": 351.0358269230769, "end": 351.19582692307694, "probability": 1.0 }, { "word": "的", "start": 351.19582692307694, "end": 351.31582692307694, "probability": 0.99658203125 }, { "word": "注意", "start": 351.31582692307694, "end": 351.55582692307695, "probability": 0.9970703125 }, { "word": "力", "start": 351.55582692307695, "end": 351.69582692307694, "probability": 0.99951171875 }, { "word": "被", "start": 351.69582692307694, "end": 351.85582692307696, "probability": 0.9970703125 }, { "word": "压", "start": 351.85582692307696, "end": 352.0358269230769, "probability": 0.9990234375 }, { "word": "缩", "start": 352.0358269230769, "end": 352.1558269230769, "probability": 0.9990234375 }, { "word": "到", "start": 352.1558269230769, "end": 352.2758269230769, "probability": 0.9951171875 }, { "word": "什么", "start": 352.2758269230769, "end": 352.5358269230769, "probability": 0.9970703125 }, { "word": "程", "start": 352.5358269230769, "end": 352.67582692307695, "probability": 0.99951171875 }, { "word": "度", "start": 352.67582692307695, "end": 352.91582692307696, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09965962779765226, "compression_ratio": 1.1941489361702127, "no_speech_prob": 2.60604888363547e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "HCA或CSA之间交错的混乱混合注意力它们的主要改动是彻底弃用HCA于是全聚分支现在只剩纯CSAR从而得到更简单的CSAR SWA混合架构但即便CSAR已经减少了跨层所需存储的KV量仍存在另一个相当巨大的低效问题因为无论预填充阶段的注意力被压缩到什么程度模型人必须先让所有这些输入Token通过网络才能生成第一个Token设想你给一个常规40层Transformer" } ], "id": 164, "review_annotation": "", "row_id": 164, "bounded_row_id": 164, "parent_row_ids": [ 164 ], "parent_row_id": 164, "packet_id": 12, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 330.25582692307694, "packet_end": 359.44381250000004, "raw_text": "因为无论预填充阶段的注意力被压缩到什么程度", "selected_text": "因为无论预填充阶段的注意力被压缩到什么程度", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 164 ], "packet_id": 12, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 352.91582692307696, "end": 355.87582692307694, "text": "模型人必须先让所有这些输入Token通过网络", "raw_asr_text": "模型人必须先让所有这些输入Token通过网络", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 12, "words": [ { "word": "模", "start": 352.91582692307696, "end": 353.19582692307694, "probability": 0.9921875 }, { "word": "型", "start": 353.19582692307694, "end": 353.35582692307696, "probability": 0.99951171875 }, { "word": "人", "start": 353.35582692307696, "end": 353.49582692307695, "probability": 0.9599609375 }, { "word": "必", "start": 353.49582692307695, "end": 353.69582692307694, "probability": 0.9970703125 }, { "word": "须", "start": 353.69582692307694, "end": 353.7758269230769, "probability": 0.99951171875 }, { "word": "先", "start": 353.7758269230769, "end": 353.91582692307696, "probability": 0.99755859375 }, { "word": "让", "start": 353.91582692307696, "end": 354.05582692307695, "probability": 0.99853515625 }, { "word": "所有", "start": 354.05582692307695, "end": 354.37582692307694, "probability": 0.99365234375 }, { "word": "这", "start": 354.37582692307694, "end": 354.51582692307693, "probability": 0.970703125 }, { "word": "些", "start": 354.51582692307693, "end": 354.63582692307693, "probability": 0.99951171875 }, { "word": "输", "start": 354.63582692307693, "end": 354.7758269230769, "probability": 0.9990234375 }, { "word": "入", "start": 354.7758269230769, "end": 354.87582692307694, "probability": 0.99853515625 }, { "word": "T", "start": 354.87582692307694, "end": 354.99582692307695, "probability": 0.525390625 }, { "word": "oken", "start": 354.99582692307695, "end": 355.1558269230769, "probability": 0.79150390625 }, { "word": "通", "start": 355.1558269230769, "end": 355.3958269230769, "probability": 0.92041015625 }, { "word": "过", "start": 355.3958269230769, "end": 355.5358269230769, "probability": 0.99951171875 }, { "word": "网", "start": 355.5358269230769, "end": 355.73582692307696, "probability": 0.99951171875 }, { "word": "络", "start": 355.73582692307696, "end": 355.87582692307694, "probability": 0.99560546875 } ], "asr_metrics": { "avg_logprob": -0.09965962779765226, "compression_ratio": 1.1941489361702127, "no_speech_prob": 2.60604888363547e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "HCA或CSA之间交错的混乱混合注意力它们的主要改动是彻底弃用HCA于是全聚分支现在只剩纯CSAR从而得到更简单的CSAR SWA混合架构但即便CSAR已经减少了跨层所需存储的KV量仍存在另一个相当巨大的低效问题因为无论预填充阶段的注意力被压缩到什么程度模型人必须先让所有这些输入Token通过网络才能生成第一个Token设想你给一个常规40层Transformer" } ], "id": 165, "review_annotation": "", "row_id": 165, "bounded_row_id": 165, "parent_row_ids": [ 165 ], "parent_row_id": 165, "packet_id": 12, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 330.25582692307694, "packet_end": 359.44381250000004, "raw_text": "模型人必须先让所有这些输入Token通过网络", "selected_text": "模型人必须先让所有这些输入Token通过网络", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 165 ], "packet_id": 12, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 355.87582692307694, "end": 357.0958269230769, "text": "才能生成第一个Token", "raw_asr_text": "才能生成第一个Token", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 12, "words": [ { "word": "才", "start": 355.87582692307694, "end": 356.17582692307695, "probability": 0.9951171875 }, { "word": "能", "start": 356.17582692307695, "end": 356.31582692307694, "probability": 0.998046875 }, { "word": "生", "start": 356.31582692307694, "end": 356.41582692307696, "probability": 0.98046875 }, { "word": "成", "start": 356.41582692307696, "end": 356.55582692307695, "probability": 0.99267578125 }, { "word": "第一", "start": 356.55582692307695, "end": 356.73582692307696, "probability": 0.94091796875 }, { "word": "个", "start": 356.73582692307696, "end": 356.85582692307696, "probability": 0.9970703125 }, { "word": "T", "start": 356.85582692307696, "end": 356.9558269230769, "probability": 0.96484375 }, { "word": "oken", "start": 356.9558269230769, "end": 357.0958269230769, "probability": 0.99609375 } ], "asr_metrics": { "avg_logprob": -0.09965962779765226, "compression_ratio": 1.1941489361702127, "no_speech_prob": 2.60604888363547e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "HCA或CSA之间交错的混乱混合注意力它们的主要改动是彻底弃用HCA于是全聚分支现在只剩纯CSAR从而得到更简单的CSAR SWA混合架构但即便CSAR已经减少了跨层所需存储的KV量仍存在另一个相当巨大的低效问题因为无论预填充阶段的注意力被压缩到什么程度模型人必须先让所有这些输入Token通过网络才能生成第一个Token设想你给一个常规40层Transformer" } ], "id": 166, "review_annotation": "", "row_id": 166, "bounded_row_id": 166, "parent_row_ids": [ 166 ], "parent_row_id": 166, "packet_id": 12, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 330.25582692307694, "packet_end": 359.44381250000004, "raw_text": "才能生成第一个Token", "selected_text": "才能生成第一个Token", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 166 ], "packet_id": 12, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 357.0958269230769, "end": 358.6438125, "text": "设想你给一个常规40层Transformer", "raw_asr_text": "设想你给一个常规40层Transformer", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 12, "words": [ { "word": "设", "start": 357.0958269230769, "end": 357.4558269230769, "probability": 0.931640625 }, { "word": "想", "start": 357.4558269230769, "end": 357.57582692307693, "probability": 0.99609375 }, { "word": "你", "start": 357.57582692307693, "end": 357.7158269230769, "probability": 0.98486328125 }, { "word": "给", "start": 357.7158269230769, "end": 357.8358269230769, "probability": 0.99853515625 }, { "word": "一个", "start": 357.8358269230769, "end": 358.01582692307693, "probability": 0.9912109375 }, { "word": "常", "start": 358.01582692307693, "end": 358.2158269230769, "probability": 0.986328125 }, { "word": "规", "start": 358.2158269230769, "end": 358.5358269230769, "probability": 0.99951171875 }, { "word": "40", "start": 358.5358269230769, "end": 358.8358269230769, "probability": 0.984375 }, { "word": "层", "start": 358.8358269230769, "end": 359.0358269230769, "probability": 0.9970703125 }, { "word": "Trans", "start": 359.0358269230769, "end": 359.19582692307694, "probability": 0.93603515625 }, { "word": "form", "start": 359.19582692307694, "end": 359.41582692307696, "probability": 0.994140625 }, { "word": "er", "start": 359.41582692307696, "end": 359.41582692307696, "probability": 0.6748046875 } ], "asr_metrics": { "avg_logprob": -0.09965962779765226, "compression_ratio": 1.1941489361702127, "no_speech_prob": 2.60604888363547e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "HCA或CSA之间交错的混乱混合注意力它们的主要改动是彻底弃用HCA于是全聚分支现在只剩纯CSAR从而得到更简单的CSAR SWA混合架构但即便CSAR已经减少了跨层所需存储的KV量仍存在另一个相当巨大的低效问题因为无论预填充阶段的注意力被压缩到什么程度模型人必须先让所有这些输入Token通过网络才能生成第一个Token设想你给一个常规40层Transformer" } ], "id": 167, "review_annotation": "", "row_id": 167, "bounded_row_id": 167, "parent_row_ids": [ 167 ], "parent_row_id": 167, "packet_id": 12, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 330.25582692307694, "packet_end": 359.44381250000004, "raw_text": "设想你给一个常规40层Transformer", "selected_text": "设想你给一个常规40层Transformer", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 167 ], "packet_id": 12, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 358.6438125, "end": 359.5238125, "text": "一个常规40层Transformer", "raw_asr_text": "一个常规40层Transformer", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "一个", "start": 357.8438125, "end": 358.00381250000004, "probability": 0.83154296875 }, { "word": "常", "start": 358.00381250000004, "end": 358.2238125, "probability": 0.97802734375 }, { "word": "规", "start": 358.2238125, "end": 358.5438125, "probability": 1.0 }, { "word": "40", "start": 358.5438125, "end": 358.82381250000003, "probability": 0.974609375 }, { "word": "层", "start": 358.82381250000003, "end": 359.0238125, "probability": 0.99365234375 }, { "word": "Trans", "start": 359.0238125, "end": 359.2038125, "probability": 0.76611328125 }, { "word": "form", "start": 359.2038125, "end": 359.38381250000003, "probability": 0.99853515625 }, { "word": "er", "start": 359.38381250000003, "end": 359.5238125, "probability": 0.98291015625 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 168, "review_annotation": "", "row_id": 168, "bounded_row_id": 168, "parent_row_ids": [ 168 ], "parent_row_id": 168, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "一个常规40层Transformer", "selected_text": "一个常规40层Transformer", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 168 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 359.5238125, "end": 361.2238125, "text": "输入100万Token的提示词", "raw_asr_text": "输入100万Token的提示词", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "输", "start": 359.5238125, "end": 359.6638125, "probability": 0.982421875 }, { "word": "入", "start": 359.6638125, "end": 359.7838125, "probability": 0.99951171875 }, { "word": "100", "start": 359.7838125, "end": 360.0238125, "probability": 0.95361328125 }, { "word": "万", "start": 360.0238125, "end": 360.2438125, "probability": 0.9638671875 }, { "word": "T", "start": 360.2438125, "end": 360.44381250000004, "probability": 0.64990234375 }, { "word": "oken", "start": 360.44381250000004, "end": 360.6038125, "probability": 0.9755859375 }, { "word": "的", "start": 360.6038125, "end": 360.7238125, "probability": 0.986328125 }, { "word": "提", "start": 360.7238125, "end": 360.8638125, "probability": 0.998046875 }, { "word": "示", "start": 360.8638125, "end": 361.00381250000004, "probability": 0.99951171875 }, { "word": "词", "start": 361.00381250000004, "end": 361.2238125, "probability": 0.9951171875 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 169, "review_annotation": "", "row_id": 169, "bounded_row_id": 169, "parent_row_ids": [ 169 ], "parent_row_id": 169, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "输入100万Token的提示词", "selected_text": "输入100万Token的提示词", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 169 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 361.2238125, "end": 362.82381250000003, "text": "全部100万个Token", "raw_asr_text": "全部100万个Token", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "全部", "start": 361.2238125, "end": 361.9238125, "probability": 0.8779296875 }, { "word": "100", "start": 361.9238125, "end": 362.2438125, "probability": 0.9931640625 }, { "word": "万", "start": 362.2438125, "end": 362.4038125, "probability": 0.9970703125 }, { "word": "个", "start": 362.4038125, "end": 362.50381250000004, "probability": 0.9453125 }, { "word": "T", "start": 362.50381250000004, "end": 362.6238125, "probability": 0.81591796875 }, { "word": "oken", "start": 362.6238125, "end": 362.82381250000003, "probability": 0.83837890625 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 170, "review_annotation": "", "row_id": 170, "bounded_row_id": 170, "parent_row_ids": [ 170 ], "parent_row_id": 170, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "全部100万个Token", "selected_text": "全部100万个Token", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 170 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 362.82381250000003, "end": 365.76381250000003, "text": "仍必须从第一层一路通过到第40层", "raw_asr_text": "仍必须从第一层一路通过到第40层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "仍", "start": 362.82381250000003, "end": 363.26381250000003, "probability": 0.9580078125 }, { "word": "必", "start": 363.26381250000003, "end": 363.44381250000004, "probability": 0.99658203125 }, { "word": "须", "start": 363.44381250000004, "end": 363.6238125, "probability": 0.998046875 }, { "word": "从", "start": 363.6238125, "end": 363.8038125, "probability": 0.998046875 }, { "word": "第一", "start": 363.8038125, "end": 364.1038125, "probability": 0.8837890625 }, { "word": "层", "start": 364.1038125, "end": 364.32381250000003, "probability": 0.9990234375 }, { "word": "一", "start": 364.32381250000003, "end": 364.4638125, "probability": 0.99853515625 }, { "word": "路", "start": 364.4638125, "end": 364.5838125, "probability": 0.9990234375 }, { "word": "通", "start": 364.5838125, "end": 364.7238125, "probability": 0.9990234375 }, { "word": "过", "start": 364.7238125, "end": 364.88381250000003, "probability": 1.0 }, { "word": "到", "start": 364.88381250000003, "end": 365.0438125, "probability": 0.99560546875 }, { "word": "第", "start": 365.0438125, "end": 365.2438125, "probability": 0.99658203125 }, { "word": "40", "start": 365.2438125, "end": 365.50381250000004, "probability": 0.9873046875 }, { "word": "层", "start": 365.50381250000004, "end": 365.76381250000003, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 171, "review_annotation": "", "row_id": 171, "bounded_row_id": 171, "parent_row_ids": [ 171 ], "parent_row_id": 171, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "仍必须从第一层一路通过到第40层", "selected_text": "仍必须从第一层一路通过到第40层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 171 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 365.76381250000003, "end": 367.0238125, "text": "对智能体操作来说", "raw_asr_text": "对智能体操作来说", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "对", "start": 365.76381250000003, "end": 365.9238125, "probability": 0.98828125 }, { "word": "智", "start": 365.9238125, "end": 366.1038125, "probability": 0.998046875 }, { "word": "能", "start": 366.1038125, "end": 366.1838125, "probability": 0.9990234375 }, { "word": "体", "start": 366.1838125, "end": 366.3438125, "probability": 0.99560546875 }, { "word": "操", "start": 366.3438125, "end": 366.5238125, "probability": 1.0 }, { "word": "作", "start": 366.5238125, "end": 366.6638125, "probability": 0.99951171875 }, { "word": "来", "start": 366.6638125, "end": 366.7838125, "probability": 0.99853515625 }, { "word": "说", "start": 366.7838125, "end": 367.0238125, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 172, "review_annotation": "", "row_id": 172, "bounded_row_id": 172, "parent_row_ids": [ 172 ], "parent_row_id": 172, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "对智能体操作来说", "selected_text": "对智能体操作来说", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 172 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 367.0238125, "end": 367.94381250000004, "text": "这相当棘手", "raw_asr_text": "这相当棘手", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "这", "start": 367.0238125, "end": 367.26381250000003, "probability": 0.9501953125 }, { "word": "相", "start": 367.26381250000003, "end": 367.44381250000004, "probability": 0.99951171875 }, { "word": "当", "start": 367.44381250000004, "end": 367.5838125, "probability": 0.99951171875 }, { "word": "棘", "start": 367.5838125, "end": 367.7838125, "probability": 0.98583984375 }, { "word": "手", "start": 367.7838125, "end": 367.94381250000004, "probability": 0.99560546875 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 173, "review_annotation": "", "row_id": 173, "bounded_row_id": 173, "parent_row_ids": [ 173 ], "parent_row_id": 173, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "这相当棘手", "selected_text": "这相当棘手", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 173 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 367.94381250000004, "end": 370.1038125, "text": "因为如今模型会不断调用工具", "raw_asr_text": "因为如今模型会不断调用工具", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "因为", "start": 367.94381250000004, "end": 368.1838125, "probability": 0.99755859375 }, { "word": "如", "start": 368.1838125, "end": 368.38381250000003, "probability": 0.99951171875 }, { "word": "今", "start": 368.38381250000003, "end": 368.56381250000004, "probability": 0.99755859375 }, { "word": "模", "start": 368.56381250000004, "end": 368.8038125, "probability": 0.9951171875 }, { "word": "型", "start": 368.8038125, "end": 368.9838125, "probability": 0.99951171875 }, { "word": "会", "start": 368.9838125, "end": 369.1838125, "probability": 0.99951171875 }, { "word": "不", "start": 369.1838125, "end": 369.3438125, "probability": 0.9990234375 }, { "word": "断", "start": 369.3438125, "end": 369.4638125, "probability": 1.0 }, { "word": "调", "start": 369.4638125, "end": 369.6238125, "probability": 0.982421875 }, { "word": "用", "start": 369.6238125, "end": 369.7438125, "probability": 0.998046875 }, { "word": "工", "start": 369.7438125, "end": 369.94381250000004, "probability": 0.99462890625 }, { "word": "具", "start": 369.94381250000004, "end": 370.1038125, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 174, "review_annotation": "", "row_id": 174, "bounded_row_id": 174, "parent_row_ids": [ 174 ], "parent_row_id": 174, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "因为如今模型会不断调用工具", "selected_text": "因为如今模型会不断调用工具", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 174 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 370.1038125, "end": 370.8638125, "text": "读取文件", "raw_asr_text": "读取文件", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "读", "start": 370.1038125, "end": 370.3438125, "probability": 0.98193359375 }, { "word": "取", "start": 370.3438125, "end": 370.4638125, "probability": 0.9990234375 }, { "word": "文", "start": 370.4638125, "end": 370.6438125, "probability": 0.99951171875 }, { "word": "件", "start": 370.6438125, "end": 370.8638125, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 175, "review_annotation": "", "row_id": 175, "bounded_row_id": 175, "parent_row_ids": [ 175 ], "parent_row_id": 175, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "读取文件", "selected_text": "读取文件", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 175 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 370.8638125, "end": 372.00381250000004, "text": "获取搜索结果", "raw_asr_text": "获取搜索结果", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "获", "start": 370.8638125, "end": 371.1438125, "probability": 0.96875 }, { "word": "取", "start": 371.1438125, "end": 371.26381250000003, "probability": 0.9990234375 }, { "word": "搜", "start": 371.26381250000003, "end": 371.44381250000004, "probability": 0.998046875 }, { "word": "索", "start": 371.44381250000004, "end": 371.5438125, "probability": 0.9990234375 }, { "word": "结", "start": 371.5438125, "end": 371.7038125, "probability": 0.99951171875 }, { "word": "果", "start": 371.7038125, "end": 372.00381250000004, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 176, "review_annotation": "", "row_id": 176, "bounded_row_id": 176, "parent_row_ids": [ 176 ], "parent_row_id": 176, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "获取搜索结果", "selected_text": "获取搜索结果", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 176 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 372.00381250000004, "end": 373.0838125, "text": "浏览代码库", "raw_asr_text": "浏览代码库", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "浏", "start": 372.00381250000004, "end": 372.2838125, "probability": 0.96240234375 }, { "word": "览", "start": 372.2838125, "end": 372.4838125, "probability": 0.9990234375 }, { "word": "代", "start": 372.4838125, "end": 372.6438125, "probability": 0.884765625 }, { "word": "码", "start": 372.6438125, "end": 372.82381250000003, "probability": 0.9990234375 }, { "word": "库", "start": 372.82381250000003, "end": 373.0838125, "probability": 0.994140625 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 177, "review_annotation": "", "row_id": 177, "bounded_row_id": 177, "parent_row_ids": [ 177 ], "parent_row_id": 177, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "浏览代码库", "selected_text": "浏览代码库", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 177 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 373.0838125, "end": 375.9238125, "text": "然后把所有这些再加回上下文理", "raw_asr_text": "然后把所有这些再加回上下文理", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "然后", "start": 373.0838125, "end": 373.4638125, "probability": 0.9912109375 }, { "word": "把", "start": 373.4638125, "end": 373.6838125, "probability": 0.998046875 }, { "word": "所有", "start": 373.6838125, "end": 374.0838125, "probability": 0.99755859375 }, { "word": "这", "start": 374.0838125, "end": 374.2238125, "probability": 0.99853515625 }, { "word": "些", "start": 374.2238125, "end": 374.3638125, "probability": 0.99951171875 }, { "word": "再", "start": 374.3638125, "end": 374.5238125, "probability": 0.99169921875 }, { "word": "加", "start": 374.5238125, "end": 374.7838125, "probability": 0.99853515625 }, { "word": "回", "start": 374.7838125, "end": 374.9838125, "probability": 0.99951171875 }, { "word": "上", "start": 374.9838125, "end": 375.2838125, "probability": 0.9990234375 }, { "word": "下", "start": 375.2838125, "end": 375.5238125, "probability": 0.99951171875 }, { "word": "文", "start": 375.5238125, "end": 375.7438125, "probability": 0.87255859375 }, { "word": "理", "start": 375.7438125, "end": 375.9238125, "probability": 0.98779296875 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 178, "review_annotation": "", "row_id": 178, "bounded_row_id": 178, "parent_row_ids": [ 178 ], "parent_row_id": 178, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "然后把所有这些再加回上下文理", "selected_text": "然后把所有这些再加回上下文理", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 178 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 375.9238125, "end": 379.2438125, "text": "这意味着模型会反复耗费多的离谱的算力", "raw_asr_text": "这意味着模型会反复耗费多的离谱的算力", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "这", "start": 375.9238125, "end": 376.1238125, "probability": 0.998046875 }, { "word": "意", "start": 376.1238125, "end": 376.3038125, "probability": 0.99853515625 }, { "word": "味", "start": 376.3038125, "end": 376.4038125, "probability": 0.9990234375 }, { "word": "着", "start": 376.4038125, "end": 376.5238125, "probability": 0.99853515625 }, { "word": "模", "start": 376.5238125, "end": 376.7238125, "probability": 0.99853515625 }, { "word": "型", "start": 376.7238125, "end": 376.8638125, "probability": 1.0 }, { "word": "会", "start": 376.8638125, "end": 377.00381250000004, "probability": 0.9990234375 }, { "word": "反", "start": 377.00381250000004, "end": 377.38381250000003, "probability": 0.99267578125 }, { "word": "复", "start": 377.38381250000003, "end": 377.7038125, "probability": 0.99853515625 }, { "word": "耗", "start": 377.7038125, "end": 377.9638125, "probability": 0.99755859375 }, { "word": "费", "start": 377.9638125, "end": 378.1438125, "probability": 0.9990234375 }, { "word": "多", "start": 378.1438125, "end": 378.3438125, "probability": 0.998046875 }, { "word": "的", "start": 378.3438125, "end": 378.4838125, "probability": 0.96484375 }, { "word": "离", "start": 378.4838125, "end": 378.6838125, "probability": 0.9931640625 }, { "word": "谱", "start": 378.6838125, "end": 378.8038125, "probability": 0.9970703125 }, { "word": "的", "start": 378.8038125, "end": 378.9238125, "probability": 0.97802734375 }, { "word": "算", "start": 378.9238125, "end": 379.06381250000004, "probability": 0.99951171875 }, { "word": "力", "start": 379.06381250000004, "end": 379.2438125, "probability": 0.97607421875 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 179, "review_annotation": "", "row_id": 179, "bounded_row_id": 179, "parent_row_ids": [ 179 ], "parent_row_id": 179, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "这意味着模型会反复耗费多的离谱的算力", "selected_text": "这意味着模型会反复耗费多的离谱的算力", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 179 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 379.2438125, "end": 380.3438125, "text": "只是阅读", "raw_asr_text": "只是阅读", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "只是", "start": 379.2438125, "end": 379.6438125, "probability": 0.99560546875 }, { "word": "阅", "start": 379.6438125, "end": 380.1038125, "probability": 0.90625 }, { "word": "读", "start": 380.1038125, "end": 380.3438125, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 180, "review_annotation": "", "row_id": 180, "bounded_row_id": 180, "parent_row_ids": [ 180 ], "parent_row_id": 180, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "只是阅读", "selected_text": "只是阅读", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 180 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 380.3438125, "end": 381.5838125, "text": "在推理场景下", "raw_asr_text": "在推理场景下", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "在", "start": 380.3438125, "end": 380.7838125, "probability": 0.97802734375 }, { "word": "推", "start": 380.7838125, "end": 381.00381250000004, "probability": 1.0 }, { "word": "理", "start": 381.00381250000004, "end": 381.1438125, "probability": 0.99951171875 }, { "word": "场", "start": 381.1438125, "end": 381.2838125, "probability": 0.99267578125 }, { "word": "景", "start": 381.2838125, "end": 381.4038125, "probability": 0.9990234375 }, { "word": "下", "start": 381.4038125, "end": 381.5838125, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 181, "review_annotation": "", "row_id": 181, "bounded_row_id": 181, "parent_row_ids": [ 181 ], "parent_row_id": 181, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "在推理场景下", "selected_text": "在推理场景下", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 181 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 381.5838125, "end": 383.94381250000004, "text": "大多数前沿实验室和推理服务商", "raw_asr_text": "大多数前沿实验室和推理服务商", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "大", "start": 381.5838125, "end": 381.8438125, "probability": 0.99755859375 }, { "word": "多", "start": 381.8438125, "end": 382.0238125, "probability": 0.9990234375 }, { "word": "数", "start": 382.0238125, "end": 382.2038125, "probability": 0.99951171875 }, { "word": "前", "start": 382.2038125, "end": 382.3638125, "probability": 0.9677734375 }, { "word": "沿", "start": 382.3638125, "end": 382.50381250000004, "probability": 0.9921875 }, { "word": "实", "start": 382.50381250000004, "end": 382.6638125, "probability": 0.9990234375 }, { "word": "验", "start": 382.6638125, "end": 382.76381250000003, "probability": 0.9931640625 }, { "word": "室", "start": 382.76381250000003, "end": 382.94381250000004, "probability": 0.9990234375 }, { "word": "和", "start": 382.94381250000004, "end": 383.1838125, "probability": 0.99951171875 }, { "word": "推", "start": 383.1838125, "end": 383.3638125, "probability": 1.0 }, { "word": "理", "start": 383.3638125, "end": 383.50381250000004, "probability": 1.0 }, { "word": "服", "start": 383.50381250000004, "end": 383.6438125, "probability": 0.9990234375 }, { "word": "务", "start": 383.6438125, "end": 383.7438125, "probability": 1.0 }, { "word": "商", "start": 383.7438125, "end": 383.94381250000004, "probability": 0.99462890625 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 182, "review_annotation": "", "row_id": 182, "bounded_row_id": 182, "parent_row_ids": [ 182 ], "parent_row_id": 182, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "大多数前沿实验室和推理服务商", "selected_text": "大多数前沿实验室和推理服务商", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 182 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 383.94381250000004, "end": 386.2317980769231, "text": "已经将预填充和解码拆分到不同的", "raw_asr_text": "已经将预填充和解码拆分到不同的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 13, "words": [ { "word": "已经", "start": 383.94381250000004, "end": 384.26381250000003, "probability": 0.99755859375 }, { "word": "将", "start": 384.26381250000003, "end": 384.56381250000004, "probability": 0.99951171875 }, { "word": "预", "start": 384.56381250000004, "end": 384.8038125, "probability": 0.98193359375 }, { "word": "填", "start": 384.8038125, "end": 385.0438125, "probability": 0.97314453125 }, { "word": "充", "start": 385.0438125, "end": 385.2238125, "probability": 0.99658203125 }, { "word": "和", "start": 385.2238125, "end": 385.4238125, "probability": 0.99853515625 }, { "word": "解", "start": 385.4238125, "end": 385.5438125, "probability": 0.994140625 }, { "word": "码", "start": 385.5438125, "end": 385.76381250000003, "probability": 0.99853515625 }, { "word": "拆", "start": 385.76381250000003, "end": 386.00381250000004, "probability": 0.9990234375 }, { "word": "分", "start": 386.00381250000004, "end": 386.1838125, "probability": 0.99951171875 }, { "word": "到", "start": 386.1838125, "end": 386.3438125, "probability": 0.9970703125 }, { "word": "不同", "start": 386.3438125, "end": 386.6638125, "probability": 0.99755859375 }, { "word": "的", "start": 386.6638125, "end": 386.9238125, "probability": 0.994140625 } ], "asr_metrics": { "avg_logprob": -0.08732545542162518, "compression_ratio": 1.2186666666666666, "no_speech_prob": 3.3856490527783833e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "一个常规40层Transformer输入100万Token的提示词全部100万个Token仍必须从第一层一路通过到第40层对智能体操作来说这相当棘手因为如今模型会不断调用工具读取文件获取搜索结果浏览代码库然后把所有这些再加回上下文理这意味着模型会反复耗费多的离谱的算力只是阅读在推理场景下大多数前沿实验室和推理服务商已经将预填充和解码拆分到不同的我看你了" } ], "id": 183, "review_annotation": "", "row_id": 183, "bounded_row_id": 183, "parent_row_ids": [ 183 ], "parent_row_id": 183, "packet_id": 13, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 357.8438125, "packet_end": 387.0317980769231, "raw_text": "已经将预填充和解码拆分到不同的", "selected_text": "已经将预填充和解码拆分到不同的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 183 ], "packet_id": 13, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 386.2317980769231, "end": 388.1517980769231, "text": "解码拆分到不同的GPU集群运行", "raw_asr_text": "解码拆分到不同的GPU集群运行", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 14, "words": [ { "word": "解", "start": 385.4317980769231, "end": 385.5317980769231, "probability": 0.476318359375 }, { "word": "码", "start": 385.5317980769231, "end": 385.77179807692306, "probability": 0.9833984375 }, { "word": "拆", "start": 385.77179807692306, "end": 386.01179807692307, "probability": 0.98876953125 }, { "word": "分", "start": 386.01179807692307, "end": 386.1717980769231, "probability": 0.99951171875 }, { "word": "到", "start": 386.1717980769231, "end": 386.33179807692306, "probability": 0.99462890625 }, { "word": "不同", "start": 386.33179807692306, "end": 386.6717980769231, "probability": 0.994140625 }, { "word": "的", "start": 386.6717980769231, "end": 386.95179807692307, "probability": 0.9921875 }, { "word": "G", "start": 386.95179807692307, "end": 387.21179807692306, "probability": 0.8408203125 }, { "word": "PU", "start": 387.21179807692306, "end": 387.4317980769231, "probability": 0.9931640625 }, { "word": "集", "start": 387.4317980769231, "end": 387.5917980769231, "probability": 0.93701171875 }, { "word": "群", "start": 387.5917980769231, "end": 387.77179807692306, "probability": 0.998046875 }, { "word": "运", "start": 387.77179807692306, "end": 387.9117980769231, "probability": 0.9951171875 }, { "word": "行", "start": 387.9117980769231, "end": 388.1517980769231, "probability": 0.99560546875 } ], "asr_metrics": { "avg_logprob": -0.07952134874131944, "compression_ratio": 1.2622950819672132, "no_speech_prob": 3.243482912806961e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "解码拆分到不同的GPU集群运行因为这两种工作负载的需求差异极大其中预填充属于算力密集型任务解码则属于内存带宽密集型任务DeepSea也采用了这一做法但在架构层面进行了调整推出了名为CED的方案全称为因果编码器解码器架构实际上CED背后的核心构想并非DeepSea首创这一构想的灵感来自2024年发表的一篇论文论文题为You only catch once简称Yoko优酷方面发现在长谷" } ], "id": 184, "review_annotation": "", "row_id": 184, "bounded_row_id": 184, "parent_row_ids": [ 184 ], "parent_row_id": 184, "packet_id": 14, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 385.4317980769231, "packet_end": 414.6197836538462, "raw_text": "解码拆分到不同的GPU集群运行", "selected_text": "解码拆分到不同的GPU集群运行", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 184 ], "packet_id": 14, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 388.1517980769231, "end": 390.7917980769231, "text": "因为这两种工作负载的需求差异极大", "raw_asr_text": "因为这两种工作负载的需求差异极大", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 14, "words": [ { "word": "因为", "start": 388.1517980769231, "end": 388.5717980769231, "probability": 0.92626953125 }, { "word": "这", "start": 388.5717980769231, "end": 388.7517980769231, "probability": 0.99560546875 }, { "word": "两", "start": 388.7517980769231, "end": 388.89179807692307, "probability": 0.99365234375 }, { "word": "种", "start": 388.89179807692307, "end": 389.0317980769231, "probability": 0.99755859375 }, { "word": "工作", "start": 389.0317980769231, "end": 389.3117980769231, "probability": 0.99462890625 }, { "word": "负", "start": 389.3117980769231, "end": 389.51179807692307, "probability": 0.98291015625 }, { "word": "载", "start": 389.51179807692307, "end": 389.6517980769231, "probability": 0.9990234375 }, { "word": "的", "start": 389.6517980769231, "end": 389.7517980769231, "probability": 0.9951171875 }, { "word": "需", "start": 389.7517980769231, "end": 389.9117980769231, "probability": 0.99658203125 }, { "word": "求", "start": 389.9117980769231, "end": 390.0917980769231, "probability": 1.0 }, { "word": "差", "start": 390.0917980769231, "end": 390.2517980769231, "probability": 0.99609375 }, { "word": "异", "start": 390.2517980769231, "end": 390.4317980769231, "probability": 0.99609375 }, { "word": "极", "start": 390.4317980769231, "end": 390.5917980769231, "probability": 0.99462890625 }, { "word": "大", "start": 390.5917980769231, "end": 390.7917980769231, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.07952134874131944, "compression_ratio": 1.2622950819672132, "no_speech_prob": 3.243482912806961e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "解码拆分到不同的GPU集群运行因为这两种工作负载的需求差异极大其中预填充属于算力密集型任务解码则属于内存带宽密集型任务DeepSea也采用了这一做法但在架构层面进行了调整推出了名为CED的方案全称为因果编码器解码器架构实际上CED背后的核心构想并非DeepSea首创这一构想的灵感来自2024年发表的一篇论文论文题为You only catch once简称Yoko优酷方面发现在长谷" } ], "id": 185, "review_annotation": "", "row_id": 185, "bounded_row_id": 185, "parent_row_ids": [ 185 ], "parent_row_id": 185, "packet_id": 14, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 385.4317980769231, "packet_end": 414.6197836538462, "raw_text": "因为这两种工作负载的需求差异极大", "selected_text": "因为这两种工作负载的需求差异极大", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 185 ], "packet_id": 14, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 390.7917980769231, "end": 393.6317980769231, "text": "其中预填充属于算力密集型任务", "raw_asr_text": "其中预填充属于算力密集型任务", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 14, "words": [ { "word": "其", "start": 390.7917980769231, "end": 391.3717980769231, "probability": 0.982421875 }, { "word": "中", "start": 391.3717980769231, "end": 391.5717980769231, "probability": 0.9990234375 }, { "word": "预", "start": 391.5717980769231, "end": 391.7917980769231, "probability": 0.70556640625 }, { "word": "填", "start": 391.7917980769231, "end": 392.01179807692307, "probability": 0.9853515625 }, { "word": "充", "start": 392.01179807692307, "end": 392.1917980769231, "probability": 0.9931640625 }, { "word": "属", "start": 392.1917980769231, "end": 392.4717980769231, "probability": 0.9970703125 }, { "word": "于", "start": 392.4717980769231, "end": 392.5917980769231, "probability": 0.99853515625 }, { "word": "算", "start": 392.5917980769231, "end": 392.7517980769231, "probability": 0.99072265625 }, { "word": "力", "start": 392.7517980769231, "end": 392.89179807692307, "probability": 0.982421875 }, { "word": "密", "start": 392.89179807692307, "end": 393.0517980769231, "probability": 0.97607421875 }, { "word": "集", "start": 393.0517980769231, "end": 393.1917980769231, "probability": 0.982421875 }, { "word": "型", "start": 393.1917980769231, "end": 393.3117980769231, "probability": 0.85986328125 }, { "word": "任", "start": 393.3117980769231, "end": 393.45179807692307, "probability": 0.9921875 }, { "word": "务", "start": 393.45179807692307, "end": 393.6317980769231, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.07952134874131944, "compression_ratio": 1.2622950819672132, "no_speech_prob": 3.243482912806961e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "解码拆分到不同的GPU集群运行因为这两种工作负载的需求差异极大其中预填充属于算力密集型任务解码则属于内存带宽密集型任务DeepSea也采用了这一做法但在架构层面进行了调整推出了名为CED的方案全称为因果编码器解码器架构实际上CED背后的核心构想并非DeepSea首创这一构想的灵感来自2024年发表的一篇论文论文题为You only catch once简称Yoko优酷方面发现在长谷" } ], "id": 186, "review_annotation": "", "row_id": 186, "bounded_row_id": 186, "parent_row_ids": [ 186 ], "parent_row_id": 186, "packet_id": 14, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 385.4317980769231, "packet_end": 414.6197836538462, "raw_text": "其中预填充属于算力密集型任务", "selected_text": "其中预填充属于算力密集型任务", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 186 ], "packet_id": 14, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 393.6317980769231, "end": 395.89179807692307, "text": "解码则属于内存带宽密集型任务", "raw_asr_text": "解码则属于内存带宽密集型任务", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 14, "words": [ { "word": "解", "start": 393.6317980769231, "end": 393.8517980769231, "probability": 0.99658203125 }, { "word": "码", "start": 393.8517980769231, "end": 394.01179807692307, "probability": 1.0 }, { "word": "则", "start": 394.01179807692307, "end": 394.1317980769231, "probability": 0.99951171875 }, { "word": "属", "start": 394.1317980769231, "end": 394.3117980769231, "probability": 0.998046875 }, { "word": "于", "start": 394.3117980769231, "end": 394.4117980769231, "probability": 0.998046875 }, { "word": "内", "start": 394.4117980769231, "end": 394.5517980769231, "probability": 0.9892578125 }, { "word": "存", "start": 394.5517980769231, "end": 394.7917980769231, "probability": 1.0 }, { "word": "带", "start": 394.7917980769231, "end": 395.01179807692307, "probability": 0.9912109375 }, { "word": "宽", "start": 395.01179807692307, "end": 395.21179807692306, "probability": 0.99853515625 }, { "word": "密", "start": 395.21179807692306, "end": 395.39179807692307, "probability": 0.99658203125 }, { "word": "集", "start": 395.39179807692307, "end": 395.51179807692307, "probability": 0.99951171875 }, { "word": "型", "start": 395.51179807692307, "end": 395.6317980769231, "probability": 0.98095703125 }, { "word": "任", "start": 395.6317980769231, "end": 395.7517980769231, "probability": 0.99755859375 }, { "word": "务", "start": 395.7517980769231, "end": 395.89179807692307, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07952134874131944, "compression_ratio": 1.2622950819672132, "no_speech_prob": 3.243482912806961e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "解码拆分到不同的GPU集群运行因为这两种工作负载的需求差异极大其中预填充属于算力密集型任务解码则属于内存带宽密集型任务DeepSea也采用了这一做法但在架构层面进行了调整推出了名为CED的方案全称为因果编码器解码器架构实际上CED背后的核心构想并非DeepSea首创这一构想的灵感来自2024年发表的一篇论文论文题为You only catch once简称Yoko优酷方面发现在长谷" } ], "id": 187, "review_annotation": "", "row_id": 187, "bounded_row_id": 187, "parent_row_ids": [ 187 ], "parent_row_id": 187, "packet_id": 14, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 385.4317980769231, "packet_end": 414.6197836538462, "raw_text": "解码则属于内存带宽密集型任务", "selected_text": "解码则属于内存带宽密集型任务", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 187 ], "packet_id": 14, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 395.89179807692307, "end": 397.4917980769231, "text": "DeepSea也采用了这一做法", "raw_asr_text": "DeepSea也采用了这一做法", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 14, "words": [ { "word": "De", "start": 395.89179807692307, "end": 396.0317980769231, "probability": 0.93505859375 }, { "word": "ep", "start": 396.0317980769231, "end": 396.1317980769231, "probability": 0.96875 }, { "word": "Se", "start": 396.1317980769231, "end": 396.2517980769231, "probability": 0.219970703125 }, { "word": "a", "start": 396.2517980769231, "end": 396.2917980769231, "probability": 0.98388671875 }, { "word": "也", "start": 396.2917980769231, "end": 396.4717980769231, "probability": 0.9716796875 }, { "word": "采", "start": 396.4717980769231, "end": 396.6717980769231, "probability": 0.99853515625 }, { "word": "用", "start": 396.6717980769231, "end": 396.8117980769231, "probability": 0.9990234375 }, { "word": "了", "start": 396.8117980769231, "end": 396.9317980769231, "probability": 0.97412109375 }, { "word": "这", "start": 396.9317980769231, "end": 397.0917980769231, "probability": 0.99462890625 }, { "word": "一", "start": 397.0917980769231, "end": 397.1917980769231, "probability": 0.97900390625 }, { "word": "做", "start": 397.1917980769231, "end": 397.3117980769231, "probability": 0.982421875 }, { "word": "法", "start": 397.3117980769231, "end": 397.4917980769231, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.07952134874131944, "compression_ratio": 1.2622950819672132, "no_speech_prob": 3.243482912806961e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "解码拆分到不同的GPU集群运行因为这两种工作负载的需求差异极大其中预填充属于算力密集型任务解码则属于内存带宽密集型任务DeepSea也采用了这一做法但在架构层面进行了调整推出了名为CED的方案全称为因果编码器解码器架构实际上CED背后的核心构想并非DeepSea首创这一构想的灵感来自2024年发表的一篇论文论文题为You only catch once简称Yoko优酷方面发现在长谷" } ], "id": 188, "review_annotation": "", "row_id": 188, "bounded_row_id": 188, "parent_row_ids": [ 188 ], "parent_row_id": 188, "packet_id": 14, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 385.4317980769231, "packet_end": 414.6197836538462, "raw_text": "DeepSea也采用了这一做法", "selected_text": "DeepSea也采用了这一做法", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 188 ], "packet_id": 14, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 397.4917980769231, "end": 399.2317980769231, "text": "但在架构层面进行了调整", "raw_asr_text": "但在架构层面进行了调整", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 14, "words": [ { "word": "但", "start": 397.4917980769231, "end": 397.6717980769231, "probability": 0.9833984375 }, { "word": "在", "start": 397.6717980769231, "end": 397.8517980769231, "probability": 0.99609375 }, { "word": "架", "start": 397.8517980769231, "end": 398.0317980769231, "probability": 0.998046875 }, { "word": "构", "start": 398.0317980769231, "end": 398.1917980769231, "probability": 0.99853515625 }, { "word": "层", "start": 398.1917980769231, "end": 398.3517980769231, "probability": 0.99755859375 }, { "word": "面", "start": 398.3517980769231, "end": 398.4917980769231, "probability": 0.99951171875 }, { "word": "进", "start": 398.4917980769231, "end": 398.6517980769231, "probability": 0.99853515625 }, { "word": "行了", "start": 398.6517980769231, "end": 398.8517980769231, "probability": 0.9853515625 }, { "word": "调", "start": 398.8517980769231, "end": 399.0717980769231, "probability": 0.99951171875 }, { "word": "整", "start": 399.0717980769231, "end": 399.2317980769231, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07952134874131944, "compression_ratio": 1.2622950819672132, "no_speech_prob": 3.243482912806961e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "解码拆分到不同的GPU集群运行因为这两种工作负载的需求差异极大其中预填充属于算力密集型任务解码则属于内存带宽密集型任务DeepSea也采用了这一做法但在架构层面进行了调整推出了名为CED的方案全称为因果编码器解码器架构实际上CED背后的核心构想并非DeepSea首创这一构想的灵感来自2024年发表的一篇论文论文题为You only catch once简称Yoko优酷方面发现在长谷" } ], "id": 189, "review_annotation": "", "row_id": 189, "bounded_row_id": 189, "parent_row_ids": [ 189 ], "parent_row_id": 189, "packet_id": 14, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 385.4317980769231, "packet_end": 414.6197836538462, "raw_text": "但在架构层面进行了调整", "selected_text": "但在架构层面进行了调整", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 189 ], "packet_id": 14, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 399.2317980769231, "end": 400.8517980769231, "text": "推出了名为CED的方案", "raw_asr_text": "推出了名为CED的方案", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 14, "words": [ { "word": "推", "start": 399.2317980769231, "end": 399.4317980769231, "probability": 0.9560546875 }, { "word": "出", "start": 399.4317980769231, "end": 399.5917980769231, "probability": 0.9990234375 }, { "word": "了", "start": 399.5917980769231, "end": 399.6917980769231, "probability": 0.990234375 }, { "word": "名", "start": 399.6917980769231, "end": 399.7917980769231, "probability": 0.99658203125 }, { "word": "为", "start": 399.7917980769231, "end": 399.9917980769231, "probability": 0.998046875 }, { "word": "C", "start": 399.9917980769231, "end": 400.2517980769231, "probability": 0.7529296875 }, { "word": "ED", "start": 400.2517980769231, "end": 400.4117980769231, "probability": 0.9951171875 }, { "word": "的", "start": 400.4117980769231, "end": 400.5917980769231, "probability": 0.97314453125 }, { "word": "方", "start": 400.5917980769231, "end": 400.71179807692306, "probability": 0.99560546875 }, { "word": "案", "start": 400.71179807692306, "end": 400.8517980769231, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.07952134874131944, "compression_ratio": 1.2622950819672132, "no_speech_prob": 3.243482912806961e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "解码拆分到不同的GPU集群运行因为这两种工作负载的需求差异极大其中预填充属于算力密集型任务解码则属于内存带宽密集型任务DeepSea也采用了这一做法但在架构层面进行了调整推出了名为CED的方案全称为因果编码器解码器架构实际上CED背后的核心构想并非DeepSea首创这一构想的灵感来自2024年发表的一篇论文论文题为You only catch once简称Yoko优酷方面发现在长谷" } ], "id": 190, "review_annotation": "", "row_id": 190, "bounded_row_id": 190, "parent_row_ids": [ 190 ], "parent_row_id": 190, "packet_id": 14, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 385.4317980769231, "packet_end": 414.6197836538462, "raw_text": "推出了名为CED的方案", "selected_text": "推出了名为CED的方案", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 190 ], "packet_id": 14, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 400.8517980769231, "end": 403.3517980769231, "text": "全称为因果编码器解码器架构", "raw_asr_text": "全称为因果编码器解码器架构", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 14, "words": [ { "word": "全", "start": 400.8517980769231, "end": 401.1117980769231, "probability": 0.9736328125 }, { "word": "称", "start": 401.1117980769231, "end": 401.3517980769231, "probability": 1.0 }, { "word": "为", "start": 401.3517980769231, "end": 401.5917980769231, "probability": 0.99169921875 }, { "word": "因", "start": 401.5917980769231, "end": 401.7517980769231, "probability": 0.97021484375 }, { "word": "果", "start": 401.7517980769231, "end": 401.95179807692307, "probability": 1.0 }, { "word": "编", "start": 401.95179807692307, "end": 402.1517980769231, "probability": 0.9951171875 }, { "word": "码", "start": 402.1517980769231, "end": 402.3517980769231, "probability": 1.0 }, { "word": "器", "start": 402.3517980769231, "end": 402.51179807692307, "probability": 0.9970703125 }, { "word": "解", "start": 402.51179807692307, "end": 402.71179807692306, "probability": 0.98095703125 }, { "word": "码", "start": 402.71179807692306, "end": 402.8517980769231, "probability": 1.0 }, { "word": "器", "start": 402.8517980769231, "end": 403.01179807692307, "probability": 0.998046875 }, { "word": "架", "start": 403.01179807692307, "end": 403.1917980769231, "probability": 0.9990234375 }, { "word": "构", "start": 403.1917980769231, "end": 403.3517980769231, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.07952134874131944, "compression_ratio": 1.2622950819672132, "no_speech_prob": 3.243482912806961e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "解码拆分到不同的GPU集群运行因为这两种工作负载的需求差异极大其中预填充属于算力密集型任务解码则属于内存带宽密集型任务DeepSea也采用了这一做法但在架构层面进行了调整推出了名为CED的方案全称为因果编码器解码器架构实际上CED背后的核心构想并非DeepSea首创这一构想的灵感来自2024年发表的一篇论文论文题为You only catch once简称Yoko优酷方面发现在长谷" } ], "id": 191, "review_annotation": "", "row_id": 191, "bounded_row_id": 191, "parent_row_ids": [ 191 ], "parent_row_id": 191, "packet_id": 14, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 385.4317980769231, "packet_end": 414.6197836538462, "raw_text": "全称为因果编码器解码器架构", "selected_text": "全称为因果编码器解码器架构", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 191 ], "packet_id": 14, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 403.3517980769231, "end": 405.5317980769231, "text": "实际上CED背后的核心构想", "raw_asr_text": "实际上CED背后的核心构想", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 14, "words": [ { "word": "实", "start": 403.3517980769231, "end": 403.6517980769231, "probability": 0.99462890625 }, { "word": "际", "start": 403.6517980769231, "end": 403.7517980769231, "probability": 1.0 }, { "word": "上", "start": 403.7517980769231, "end": 403.89179807692307, "probability": 0.9990234375 }, { "word": "C", "start": 403.89179807692307, "end": 404.0917980769231, "probability": 0.76953125 }, { "word": "ED", "start": 404.0917980769231, "end": 404.2917980769231, "probability": 1.0 }, { "word": "背", "start": 404.2917980769231, "end": 404.5317980769231, "probability": 0.990234375 }, { "word": "后", "start": 404.5317980769231, "end": 404.6917980769231, "probability": 0.99853515625 }, { "word": "的", "start": 404.6917980769231, "end": 404.8117980769231, "probability": 0.9970703125 }, { "word": "核", "start": 404.8117980769231, "end": 405.0317980769231, "probability": 0.99951171875 }, { "word": "心", "start": 405.0317980769231, "end": 405.1917980769231, "probability": 0.99951171875 }, { "word": "构", "start": 405.1917980769231, "end": 405.3717980769231, "probability": 0.9970703125 }, { "word": "想", "start": 405.3717980769231, "end": 405.5317980769231, "probability": 0.9873046875 } ], "asr_metrics": { "avg_logprob": -0.07952134874131944, "compression_ratio": 1.2622950819672132, "no_speech_prob": 3.243482912806961e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "解码拆分到不同的GPU集群运行因为这两种工作负载的需求差异极大其中预填充属于算力密集型任务解码则属于内存带宽密集型任务DeepSea也采用了这一做法但在架构层面进行了调整推出了名为CED的方案全称为因果编码器解码器架构实际上CED背后的核心构想并非DeepSea首创这一构想的灵感来自2024年发表的一篇论文论文题为You only catch once简称Yoko优酷方面发现在长谷" } ], "id": 192, "review_annotation": "", "row_id": 192, "bounded_row_id": 192, "parent_row_ids": [ 192 ], "parent_row_id": 192, "packet_id": 14, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 385.4317980769231, "packet_end": 414.6197836538462, "raw_text": "实际上CED背后的核心构想", "selected_text": "实际上CED背后的核心构想", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 192 ], "packet_id": 14, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 405.5317980769231, "end": 406.7517980769231, "text": "并非DeepSea首创", "raw_asr_text": "并非DeepSea首创", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 14, "words": [ { "word": "并", "start": 405.5317980769231, "end": 405.7517980769231, "probability": 0.99951171875 }, { "word": "非", "start": 405.7517980769231, "end": 405.89179807692307, "probability": 0.99755859375 }, { "word": "De", "start": 405.89179807692307, "end": 406.0517980769231, "probability": 0.982421875 }, { "word": "ep", "start": 406.0517980769231, "end": 406.1317980769231, "probability": 0.99853515625 }, { "word": "Se", "start": 406.1317980769231, "end": 406.2317980769231, "probability": 0.97802734375 }, { "word": "a", "start": 406.2317980769231, "end": 406.2917980769231, "probability": 0.98388671875 }, { "word": "首", "start": 406.2917980769231, "end": 406.4917980769231, "probability": 0.658203125 }, { "word": "创", "start": 406.4917980769231, "end": 406.7517980769231, "probability": 0.99609375 } ], "asr_metrics": { "avg_logprob": -0.07952134874131944, "compression_ratio": 1.2622950819672132, "no_speech_prob": 3.243482912806961e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "解码拆分到不同的GPU集群运行因为这两种工作负载的需求差异极大其中预填充属于算力密集型任务解码则属于内存带宽密集型任务DeepSea也采用了这一做法但在架构层面进行了调整推出了名为CED的方案全称为因果编码器解码器架构实际上CED背后的核心构想并非DeepSea首创这一构想的灵感来自2024年发表的一篇论文论文题为You only catch once简称Yoko优酷方面发现在长谷" } ], "id": 193, "review_annotation": "", "row_id": 193, "bounded_row_id": 193, "parent_row_ids": [ 193 ], "parent_row_id": 193, "packet_id": 14, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 385.4317980769231, "packet_end": 414.6197836538462, "raw_text": "并非DeepSea首创", "selected_text": "并非DeepSea首创", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 193 ], "packet_id": 14, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 406.7517980769231, "end": 410.1517980769231, "text": "这一构想的灵感来自2024年发表的一篇论文", "raw_asr_text": "这一构想的灵感来自2024年发表的一篇论文", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 14, "words": [ { "word": "这", "start": 406.7517980769231, "end": 407.0317980769231, "probability": 0.99365234375 }, { "word": "一", "start": 407.0317980769231, "end": 407.1317980769231, "probability": 0.99462890625 }, { "word": "构", "start": 407.1317980769231, "end": 407.2917980769231, "probability": 0.99951171875 }, { "word": "想", "start": 407.2917980769231, "end": 407.45179807692307, "probability": 0.99658203125 }, { "word": "的", "start": 407.45179807692307, "end": 407.5517980769231, "probability": 0.9990234375 }, { "word": "灵", "start": 407.5517980769231, "end": 407.77179807692306, "probability": 0.9970703125 }, { "word": "感", "start": 407.77179807692306, "end": 407.9117980769231, "probability": 0.99609375 }, { "word": "来", "start": 407.9117980769231, "end": 408.0917980769231, "probability": 0.9921875 }, { "word": "自", "start": 408.0917980769231, "end": 408.21179807692306, "probability": 0.99951171875 }, { "word": "20", "start": 408.21179807692306, "end": 408.4317980769231, "probability": 0.99560546875 }, { "word": "24", "start": 408.4317980769231, "end": 408.8517980769231, "probability": 0.99560546875 }, { "word": "年", "start": 408.8517980769231, "end": 409.0717980769231, "probability": 0.9931640625 }, { "word": "发", "start": 409.0717980769231, "end": 409.2917980769231, "probability": 0.9951171875 }, { "word": "表", "start": 409.2917980769231, "end": 409.45179807692307, "probability": 0.99462890625 }, { "word": "的", "start": 409.45179807692307, "end": 409.5517980769231, "probability": 0.98974609375 }, { "word": "一", "start": 409.5517980769231, "end": 409.6317980769231, "probability": 0.998046875 }, { "word": "篇", "start": 409.6317980769231, "end": 409.7917980769231, "probability": 0.9990234375 }, { "word": "论", "start": 409.7917980769231, "end": 409.95179807692307, "probability": 0.998046875 }, { "word": "文", "start": 409.95179807692307, "end": 410.1517980769231, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.07952134874131944, "compression_ratio": 1.2622950819672132, "no_speech_prob": 3.243482912806961e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "解码拆分到不同的GPU集群运行因为这两种工作负载的需求差异极大其中预填充属于算力密集型任务解码则属于内存带宽密集型任务DeepSea也采用了这一做法但在架构层面进行了调整推出了名为CED的方案全称为因果编码器解码器架构实际上CED背后的核心构想并非DeepSea首创这一构想的灵感来自2024年发表的一篇论文论文题为You only catch once简称Yoko优酷方面发现在长谷" } ], "id": 194, "review_annotation": "", "row_id": 194, "bounded_row_id": 194, "parent_row_ids": [ 194 ], "parent_row_id": 194, "packet_id": 14, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 385.4317980769231, "packet_end": 414.6197836538462, "raw_text": "这一构想的灵感来自2024年发表的一篇论文", "selected_text": "这一构想的灵感来自2024年发表的一篇论文", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 194 ], "packet_id": 14, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 410.1517980769231, "end": 412.0917980769231, "text": "论文题为You only catch once", "raw_asr_text": "论文题为You only catch once", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 14, "words": [ { "word": "论", "start": 410.1517980769231, "end": 410.4717980769231, "probability": 0.9697265625 }, { "word": "文", "start": 410.4717980769231, "end": 410.6317980769231, "probability": 1.0 }, { "word": "题", "start": 410.6317980769231, "end": 410.7317980769231, "probability": 0.77392578125 }, { "word": "为", "start": 410.7317980769231, "end": 410.9117980769231, "probability": 0.9931640625 }, { "word": "You", "start": 410.9117980769231, "end": 411.0717980769231, "probability": 0.76025390625 }, { "word": " only", "start": 411.0717980769231, "end": 411.39179807692307, "probability": 0.638671875 }, { "word": " catch", "start": 411.39179807692307, "end": 411.6917980769231, "probability": 0.71875 }, { "word": " once", "start": 411.6917980769231, "end": 412.0917980769231, "probability": 0.966796875 } ], "asr_metrics": { "avg_logprob": -0.07952134874131944, "compression_ratio": 1.2622950819672132, "no_speech_prob": 3.243482912806961e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "解码拆分到不同的GPU集群运行因为这两种工作负载的需求差异极大其中预填充属于算力密集型任务解码则属于内存带宽密集型任务DeepSea也采用了这一做法但在架构层面进行了调整推出了名为CED的方案全称为因果编码器解码器架构实际上CED背后的核心构想并非DeepSea首创这一构想的灵感来自2024年发表的一篇论文论文题为You only catch once简称Yoko优酷方面发现在长谷" } ], "id": 195, "review_annotation": "", "row_id": 195, "bounded_row_id": 195, "parent_row_ids": [ 195 ], "parent_row_id": 195, "packet_id": 14, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 385.4317980769231, "packet_end": 414.6197836538462, "raw_text": "论文题为You only catch once", "selected_text": "论文题为You only catch once", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 195 ], "packet_id": 14, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 412.0917980769231, "end": 412.9917980769231, "text": "简称Yoko", "raw_asr_text": "简称Yoko", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 14, "words": [ { "word": "简", "start": 412.0917980769231, "end": 412.45179807692307, "probability": 0.9150390625 }, { "word": "称", "start": 412.45179807692307, "end": 412.6517980769231, "probability": 0.99951171875 }, { "word": "Y", "start": 412.6517980769231, "end": 412.83179807692306, "probability": 0.73681640625 }, { "word": "oko", "start": 412.83179807692306, "end": 412.9917980769231, "probability": 0.68212890625 } ], "asr_metrics": { "avg_logprob": -0.07952134874131944, "compression_ratio": 1.2622950819672132, "no_speech_prob": 3.243482912806961e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "解码拆分到不同的GPU集群运行因为这两种工作负载的需求差异极大其中预填充属于算力密集型任务解码则属于内存带宽密集型任务DeepSea也采用了这一做法但在架构层面进行了调整推出了名为CED的方案全称为因果编码器解码器架构实际上CED背后的核心构想并非DeepSea首创这一构想的灵感来自2024年发表的一篇论文论文题为You only catch once简称Yoko优酷方面发现在长谷" } ], "id": 196, "review_annotation": "", "row_id": 196, "bounded_row_id": 196, "parent_row_ids": [ 196 ], "parent_row_id": 196, "packet_id": 14, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 385.4317980769231, "packet_end": 414.6197836538462, "raw_text": "简称Yoko", "selected_text": "简称Yoko", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 196 ], "packet_id": 14, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 412.9917980769231, "end": 413.8197836538462, "text": "优酷方面发现", "raw_asr_text": "优酷方面发现", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 14, "words": [ { "word": "优", "start": 412.9917980769231, "end": 413.2317980769231, "probability": 0.923828125 }, { "word": "酷", "start": 413.2317980769231, "end": 413.3517980769231, "probability": 0.6884765625 }, { "word": "方面", "start": 413.3517980769231, "end": 413.51179807692307, "probability": 0.990234375 }, { "word": "发", "start": 413.51179807692307, "end": 413.77179807692306, "probability": 0.990234375 }, { "word": "现", "start": 413.77179807692306, "end": 413.9717980769231, "probability": 0.62646484375 } ], "asr_metrics": { "avg_logprob": -0.5440929532051086, "compression_ratio": 0.7105263157894737, "no_speech_prob": 8.014673993916865e-12 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "解码拆分到不同的GPU集群运行因为这两种工作负载的需求差异极大其中预填充属于算力密集型任务解码则属于内存带宽密集型任务DeepSea也采用了这一做法但在架构层面进行了调整推出了名为CED的方案全称为因果编码器解码器架构实际上CED背后的核心构想并非DeepSea首创这一构想的灵感来自2024年发表的一篇论文论文题为You only catch once简称Yoko优酷方面发现在长谷" } ], "id": 197, "review_annotation": "", "row_id": 197, "bounded_row_id": 197, "parent_row_ids": [ 197 ], "parent_row_id": 197, "packet_id": 14, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 385.4317980769231, "packet_end": 414.6197836538462, "raw_text": "优酷方面发现", "selected_text": "优酷方面发现", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 197 ], "packet_id": 14, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 413.8197836538462, "end": 415.59978365384615, "text": "Yoko方面发现在常规Transformer架构中", "raw_asr_text": "Yoko方面发现在常规Transformer架构中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 15, "words": [ { "word": "Y", "start": 413.01978365384616, "end": 413.21978365384615, "probability": 0.6865234375 }, { "word": "oko", "start": 413.21978365384615, "end": 413.3197836538462, "probability": 0.8125 }, { "word": "方面", "start": 413.3197836538462, "end": 413.53978365384614, "probability": 0.97509765625 }, { "word": "发", "start": 413.53978365384614, "end": 413.77978365384615, "probability": 0.998046875 }, { "word": "现在", "start": 413.77978365384615, "end": 414.1197836538462, "probability": 0.62353515625 }, { "word": "常", "start": 414.1197836538462, "end": 414.53978365384614, "probability": 0.9990234375 }, { "word": "规", "start": 414.53978365384614, "end": 414.69978365384617, "probability": 0.9990234375 }, { "word": "Trans", "start": 414.69978365384617, "end": 414.83978365384615, "probability": 0.8681640625 }, { "word": "form", "start": 414.83978365384615, "end": 415.01978365384616, "probability": 0.99853515625 }, { "word": "er", "start": 415.01978365384616, "end": 415.13978365384617, "probability": 0.9892578125 }, { "word": "架", "start": 415.13978365384617, "end": 415.27978365384615, "probability": 0.99609375 }, { "word": "构", "start": 415.27978365384615, "end": 415.41978365384614, "probability": 0.9990234375 }, { "word": "中", "start": 415.41978365384614, "end": 415.59978365384615, "probability": 0.98828125 } ], "asr_metrics": { "avg_logprob": -0.05282413679669161, "compression_ratio": 1.3504043126684635, "no_speech_prob": 1.997897333683163e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "Yoko方面发现在常规Transformer架构中每一层都会生成并存储专属的全局KB缓存所以如果模型有40层你实际上要保存40份不同版本的记忆数据为了解决这种重复存储的问题他们提出了名为解码器解码器架构的解决方案模型的下半部分是自解码器通过高效的因果自注意力机制处理输入序列生成提示词的上下文化表示随后Yoko基于下半部分的输出生成共享的全局KB缓存模型上半部分则为交叉" } ], "id": 198, "review_annotation": "", "row_id": 198, "bounded_row_id": 198, "parent_row_ids": [ 198 ], "parent_row_id": 198, "packet_id": 15, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 413.01978365384616, "packet_end": 442.20776923076926, "raw_text": "Yoko方面发现在常规Transformer架构中", "selected_text": "Yoko方面发现在常规Transformer架构中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 198 ], "packet_id": 15, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 415.59978365384615, "end": 418.89978365384616, "text": "每一层都会生成并存储专属的全局KB缓存", "raw_asr_text": "每一层都会生成并存储专属的全局KB缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 15, "words": [ { "word": "每", "start": 415.59978365384615, "end": 415.9397836538462, "probability": 0.9814453125 }, { "word": "一", "start": 415.9397836538462, "end": 416.03978365384614, "probability": 0.98486328125 }, { "word": "层", "start": 416.03978365384614, "end": 416.15978365384615, "probability": 1.0 }, { "word": "都", "start": 416.15978365384615, "end": 416.27978365384615, "probability": 0.9990234375 }, { "word": "会", "start": 416.27978365384615, "end": 416.4397836538462, "probability": 0.99951171875 }, { "word": "生", "start": 416.4397836538462, "end": 416.57978365384616, "probability": 0.99853515625 }, { "word": "成", "start": 416.57978365384616, "end": 416.77978365384615, "probability": 0.99462890625 }, { "word": "并", "start": 416.77978365384615, "end": 417.09978365384615, "probability": 0.99658203125 }, { "word": "存", "start": 417.09978365384615, "end": 417.33978365384615, "probability": 0.99951171875 }, { "word": "储", "start": 417.33978365384615, "end": 417.4997836538462, "probability": 0.9990234375 }, { "word": "专", "start": 417.4997836538462, "end": 417.6797836538462, "probability": 0.9970703125 }, { "word": "属", "start": 417.6797836538462, "end": 417.85978365384614, "probability": 0.998046875 }, { "word": "的", "start": 417.85978365384614, "end": 417.97978365384614, "probability": 0.9990234375 }, { "word": "全", "start": 417.97978365384614, "end": 418.15978365384615, "probability": 0.998046875 }, { "word": "局", "start": 418.15978365384615, "end": 418.25978365384617, "probability": 0.94482421875 }, { "word": "K", "start": 418.25978365384617, "end": 418.39978365384616, "probability": 0.8974609375 }, { "word": "B", "start": 418.39978365384616, "end": 418.51978365384616, "probability": 0.67578125 }, { "word": "缓", "start": 418.51978365384616, "end": 418.6797836538462, "probability": 0.99755859375 }, { "word": "存", "start": 418.6797836538462, "end": 418.89978365384616, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.05282413679669161, "compression_ratio": 1.3504043126684635, "no_speech_prob": 1.997897333683163e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "Yoko方面发现在常规Transformer架构中每一层都会生成并存储专属的全局KB缓存所以如果模型有40层你实际上要保存40份不同版本的记忆数据为了解决这种重复存储的问题他们提出了名为解码器解码器架构的解决方案模型的下半部分是自解码器通过高效的因果自注意力机制处理输入序列生成提示词的上下文化表示随后Yoko基于下半部分的输出生成共享的全局KB缓存模型上半部分则为交叉" } ], "id": 199, "review_annotation": "", "row_id": 199, "bounded_row_id": 199, "parent_row_ids": [ 199 ], "parent_row_id": 199, "packet_id": 15, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 413.01978365384616, "packet_end": 442.20776923076926, "raw_text": "每一层都会生成并存储专属的全局KB缓存", "selected_text": "每一层都会生成并存储专属的全局KB缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 199 ], "packet_id": 15, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 418.89978365384616, "end": 420.57978365384616, "text": "所以如果模型有40层", "raw_asr_text": "所以如果模型有40层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 15, "words": [ { "word": "所以", "start": 418.89978365384616, "end": 419.35978365384614, "probability": 0.99072265625 }, { "word": "如果", "start": 419.35978365384614, "end": 419.6197836538462, "probability": 0.9990234375 }, { "word": "模", "start": 419.6197836538462, "end": 419.79978365384613, "probability": 0.99951171875 }, { "word": "型", "start": 419.79978365384613, "end": 419.9397836538462, "probability": 1.0 }, { "word": "有", "start": 419.9397836538462, "end": 420.07978365384616, "probability": 1.0 }, { "word": "40", "start": 420.07978365384616, "end": 420.33978365384615, "probability": 0.99169921875 }, { "word": "层", "start": 420.33978365384615, "end": 420.57978365384616, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.05282413679669161, "compression_ratio": 1.3504043126684635, "no_speech_prob": 1.997897333683163e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "Yoko方面发现在常规Transformer架构中每一层都会生成并存储专属的全局KB缓存所以如果模型有40层你实际上要保存40份不同版本的记忆数据为了解决这种重复存储的问题他们提出了名为解码器解码器架构的解决方案模型的下半部分是自解码器通过高效的因果自注意力机制处理输入序列生成提示词的上下文化表示随后Yoko基于下半部分的输出生成共享的全局KB缓存模型上半部分则为交叉" } ], "id": 200, "review_annotation": "", "row_id": 200, "bounded_row_id": 200, "parent_row_ids": [ 200 ], "parent_row_id": 200, "packet_id": 15, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 413.01978365384616, "packet_end": 442.20776923076926, "raw_text": "所以如果模型有40层", "selected_text": "所以如果模型有40层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 200 ], "packet_id": 15, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 420.57978365384616, "end": 423.35978365384614, "text": "你实际上要保存40份不同版本的记忆数据", "raw_asr_text": "你实际上要保存40份不同版本的记忆数据", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 15, "words": [ { "word": "你", "start": 420.57978365384616, "end": 420.77978365384615, "probability": 0.994140625 }, { "word": "实", "start": 420.77978365384615, "end": 420.9397836538462, "probability": 1.0 }, { "word": "际", "start": 420.9397836538462, "end": 421.0597836538462, "probability": 1.0 }, { "word": "上", "start": 421.0597836538462, "end": 421.1797836538462, "probability": 0.99951171875 }, { "word": "要", "start": 421.1797836538462, "end": 421.29978365384613, "probability": 0.99951171875 }, { "word": "保", "start": 421.29978365384613, "end": 421.45978365384616, "probability": 0.99951171875 }, { "word": "存", "start": 421.45978365384616, "end": 421.6197836538462, "probability": 0.99951171875 }, { "word": "40", "start": 421.6197836538462, "end": 421.85978365384614, "probability": 0.99853515625 }, { "word": "份", "start": 421.85978365384614, "end": 422.0597836538462, "probability": 0.9921875 }, { "word": "不同", "start": 422.0597836538462, "end": 422.33978365384615, "probability": 0.9990234375 }, { "word": "版", "start": 422.33978365384615, "end": 422.51978365384616, "probability": 0.99951171875 }, { "word": "本", "start": 422.51978365384616, "end": 422.71978365384615, "probability": 0.99951171875 }, { "word": "的", "start": 422.71978365384615, "end": 422.8197836538462, "probability": 1.0 }, { "word": "记", "start": 422.8197836538462, "end": 422.97978365384614, "probability": 0.9921875 }, { "word": "忆", "start": 422.97978365384614, "end": 423.07978365384616, "probability": 0.9970703125 }, { "word": "数", "start": 423.07978365384616, "end": 423.19978365384617, "probability": 0.99755859375 }, { "word": "据", "start": 423.19978365384617, "end": 423.35978365384614, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.05282413679669161, "compression_ratio": 1.3504043126684635, "no_speech_prob": 1.997897333683163e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "Yoko方面发现在常规Transformer架构中每一层都会生成并存储专属的全局KB缓存所以如果模型有40层你实际上要保存40份不同版本的记忆数据为了解决这种重复存储的问题他们提出了名为解码器解码器架构的解决方案模型的下半部分是自解码器通过高效的因果自注意力机制处理输入序列生成提示词的上下文化表示随后Yoko基于下半部分的输出生成共享的全局KB缓存模型上半部分则为交叉" } ], "id": 201, "review_annotation": "", "row_id": 201, "bounded_row_id": 201, "parent_row_ids": [ 201 ], "parent_row_id": 201, "packet_id": 15, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 413.01978365384616, "packet_end": 442.20776923076926, "raw_text": "你实际上要保存40份不同版本的记忆数据", "selected_text": "你实际上要保存40份不同版本的记忆数据", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 201 ], "packet_id": 15, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 423.35978365384614, "end": 425.35978365384614, "text": "为了解决这种重复存储的问题", "raw_asr_text": "为了解决这种重复存储的问题", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 15, "words": [ { "word": "为", "start": 423.35978365384614, "end": 423.63978365384617, "probability": 0.998046875 }, { "word": "了", "start": 423.63978365384617, "end": 423.77978365384615, "probability": 1.0 }, { "word": "解", "start": 423.77978365384615, "end": 423.91978365384614, "probability": 1.0 }, { "word": "决", "start": 423.91978365384614, "end": 424.0597836538462, "probability": 1.0 }, { "word": "这", "start": 424.0597836538462, "end": 424.1797836538462, "probability": 0.9990234375 }, { "word": "种", "start": 424.1797836538462, "end": 424.3197836538462, "probability": 1.0 }, { "word": "重", "start": 424.3197836538462, "end": 424.4997836538462, "probability": 0.99951171875 }, { "word": "复", "start": 424.4997836538462, "end": 424.69978365384617, "probability": 0.9990234375 }, { "word": "存", "start": 424.69978365384617, "end": 424.89978365384616, "probability": 0.998046875 }, { "word": "储", "start": 424.89978365384616, "end": 424.9997836538462, "probability": 1.0 }, { "word": "的", "start": 424.9997836538462, "end": 425.13978365384617, "probability": 0.99951171875 }, { "word": "问题", "start": 425.13978365384617, "end": 425.35978365384614, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.05282413679669161, "compression_ratio": 1.3504043126684635, "no_speech_prob": 1.997897333683163e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "Yoko方面发现在常规Transformer架构中每一层都会生成并存储专属的全局KB缓存所以如果模型有40层你实际上要保存40份不同版本的记忆数据为了解决这种重复存储的问题他们提出了名为解码器解码器架构的解决方案模型的下半部分是自解码器通过高效的因果自注意力机制处理输入序列生成提示词的上下文化表示随后Yoko基于下半部分的输出生成共享的全局KB缓存模型上半部分则为交叉" } ], "id": 202, "review_annotation": "", "row_id": 202, "bounded_row_id": 202, "parent_row_ids": [ 202 ], "parent_row_id": 202, "packet_id": 15, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 413.01978365384616, "packet_end": 442.20776923076926, "raw_text": "为了解决这种重复存储的问题", "selected_text": "为了解决这种重复存储的问题", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 202 ], "packet_id": 15, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 425.35978365384614, "end": 428.27978365384615, "text": "他们提出了名为解码器解码器架构的解决方案", "raw_asr_text": "他们提出了名为解码器解码器架构的解决方案", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 15, "words": [ { "word": "他们", "start": 425.35978365384614, "end": 425.6197836538462, "probability": 0.96044921875 }, { "word": "提", "start": 425.6197836538462, "end": 425.8797836538462, "probability": 1.0 }, { "word": "出", "start": 425.8797836538462, "end": 426.03978365384614, "probability": 1.0 }, { "word": "了", "start": 426.03978365384614, "end": 426.13978365384617, "probability": 0.99951171875 }, { "word": "名", "start": 426.13978365384617, "end": 426.2397836538462, "probability": 0.9970703125 }, { "word": "为", "start": 426.2397836538462, "end": 426.41978365384614, "probability": 0.9990234375 }, { "word": "解", "start": 426.41978365384614, "end": 426.57978365384616, "probability": 0.98095703125 }, { "word": "码", "start": 426.57978365384616, "end": 426.75978365384617, "probability": 0.9970703125 }, { "word": "器", "start": 426.75978365384617, "end": 426.89978365384616, "probability": 0.994140625 }, { "word": "解", "start": 426.89978365384616, "end": 427.07978365384616, "probability": 0.98388671875 }, { "word": "码", "start": 427.07978365384616, "end": 427.21978365384615, "probability": 1.0 }, { "word": "器", "start": 427.21978365384615, "end": 427.33978365384615, "probability": 1.0 }, { "word": "架", "start": 427.33978365384615, "end": 427.4997836538462, "probability": 0.9970703125 }, { "word": "构", "start": 427.4997836538462, "end": 427.6197836538462, "probability": 0.99951171875 }, { "word": "的", "start": 427.6197836538462, "end": 427.71978365384615, "probability": 0.99951171875 }, { "word": "解", "start": 427.71978365384615, "end": 427.8797836538462, "probability": 1.0 }, { "word": "决", "start": 427.8797836538462, "end": 427.97978365384614, "probability": 1.0 }, { "word": "方", "start": 427.97978365384614, "end": 428.1197836538462, "probability": 1.0 }, { "word": "案", "start": 428.1197836538462, "end": 428.27978365384615, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.05282413679669161, "compression_ratio": 1.3504043126684635, "no_speech_prob": 1.997897333683163e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "Yoko方面发现在常规Transformer架构中每一层都会生成并存储专属的全局KB缓存所以如果模型有40层你实际上要保存40份不同版本的记忆数据为了解决这种重复存储的问题他们提出了名为解码器解码器架构的解决方案模型的下半部分是自解码器通过高效的因果自注意力机制处理输入序列生成提示词的上下文化表示随后Yoko基于下半部分的输出生成共享的全局KB缓存模型上半部分则为交叉" } ], "id": 203, "review_annotation": "", "row_id": 203, "bounded_row_id": 203, "parent_row_ids": [ 203 ], "parent_row_id": 203, "packet_id": 15, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 413.01978365384616, "packet_end": 442.20776923076926, "raw_text": "他们提出了名为解码器解码器架构的解决方案", "selected_text": "他们提出了名为解码器解码器架构的解决方案", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 203 ], "packet_id": 15, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 428.27978365384615, "end": 430.57978365384616, "text": "模型的下半部分是自解码器", "raw_asr_text": "模型的下半部分是自解码器", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 15, "words": [ { "word": "模", "start": 428.27978365384615, "end": 428.63978365384617, "probability": 0.99951171875 }, { "word": "型", "start": 428.63978365384617, "end": 428.8197836538462, "probability": 1.0 }, { "word": "的", "start": 428.8197836538462, "end": 428.91978365384614, "probability": 1.0 }, { "word": "下", "start": 428.91978365384614, "end": 429.09978365384615, "probability": 1.0 }, { "word": "半", "start": 429.09978365384615, "end": 429.25978365384617, "probability": 1.0 }, { "word": "部分", "start": 429.25978365384617, "end": 429.51978365384616, "probability": 0.9990234375 }, { "word": "是", "start": 429.51978365384616, "end": 429.79978365384613, "probability": 0.99951171875 }, { "word": "自", "start": 429.79978365384613, "end": 430.01978365384616, "probability": 0.978515625 }, { "word": "解", "start": 430.01978365384616, "end": 430.1797836538462, "probability": 0.99853515625 }, { "word": "码", "start": 430.1797836538462, "end": 430.3797836538462, "probability": 1.0 }, { "word": "器", "start": 430.3797836538462, "end": 430.57978365384616, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.05282413679669161, "compression_ratio": 1.3504043126684635, "no_speech_prob": 1.997897333683163e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "Yoko方面发现在常规Transformer架构中每一层都会生成并存储专属的全局KB缓存所以如果模型有40层你实际上要保存40份不同版本的记忆数据为了解决这种重复存储的问题他们提出了名为解码器解码器架构的解决方案模型的下半部分是自解码器通过高效的因果自注意力机制处理输入序列生成提示词的上下文化表示随后Yoko基于下半部分的输出生成共享的全局KB缓存模型上半部分则为交叉" } ], "id": 204, "review_annotation": "", "row_id": 204, "bounded_row_id": 204, "parent_row_ids": [ 204 ], "parent_row_id": 204, "packet_id": 15, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 413.01978365384616, "packet_end": 442.20776923076926, "raw_text": "模型的下半部分是自解码器", "selected_text": "模型的下半部分是自解码器", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 204 ], "packet_id": 15, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 430.57978365384616, "end": 434.03978365384614, "text": "通过高效的因果自注意力机制处理输入序列", "raw_asr_text": "通过高效的因果自注意力机制处理输入序列", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 15, "words": [ { "word": "通", "start": 430.57978365384616, "end": 430.83978365384615, "probability": 0.9931640625 }, { "word": "过", "start": 430.83978365384615, "end": 430.9997836538462, "probability": 0.9990234375 }, { "word": "高", "start": 430.9997836538462, "end": 431.19978365384617, "probability": 1.0 }, { "word": "效", "start": 431.19978365384617, "end": 431.35978365384614, "probability": 0.99462890625 }, { "word": "的", "start": 431.35978365384614, "end": 431.47978365384614, "probability": 0.9990234375 }, { "word": "因", "start": 431.47978365384614, "end": 431.6797836538462, "probability": 0.99658203125 }, { "word": "果", "start": 431.6797836538462, "end": 431.85978365384614, "probability": 1.0 }, { "word": "自", "start": 431.85978365384614, "end": 432.03978365384614, "probability": 0.9931640625 }, { "word": "注意", "start": 432.03978365384614, "end": 432.25978365384617, "probability": 0.994140625 }, { "word": "力", "start": 432.25978365384617, "end": 432.4397836538462, "probability": 1.0 }, { "word": "机", "start": 432.4397836538462, "end": 432.63978365384617, "probability": 0.99853515625 }, { "word": "制", "start": 432.63978365384617, "end": 432.8197836538462, "probability": 0.982421875 }, { "word": "处", "start": 432.8197836538462, "end": 433.13978365384617, "probability": 0.9990234375 }, { "word": "理", "start": 433.13978365384617, "end": 433.3197836538462, "probability": 1.0 }, { "word": "输", "start": 433.3197836538462, "end": 433.51978365384616, "probability": 0.99609375 }, { "word": "入", "start": 433.51978365384616, "end": 433.63978365384617, "probability": 1.0 }, { "word": "序", "start": 433.63978365384617, "end": 433.83978365384615, "probability": 0.998046875 }, { "word": "列", "start": 433.83978365384615, "end": 434.03978365384614, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.05282413679669161, "compression_ratio": 1.3504043126684635, "no_speech_prob": 1.997897333683163e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "Yoko方面发现在常规Transformer架构中每一层都会生成并存储专属的全局KB缓存所以如果模型有40层你实际上要保存40份不同版本的记忆数据为了解决这种重复存储的问题他们提出了名为解码器解码器架构的解决方案模型的下半部分是自解码器通过高效的因果自注意力机制处理输入序列生成提示词的上下文化表示随后Yoko基于下半部分的输出生成共享的全局KB缓存模型上半部分则为交叉" } ], "id": 205, "review_annotation": "", "row_id": 205, "bounded_row_id": 205, "parent_row_ids": [ 205 ], "parent_row_id": 205, "packet_id": 15, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 413.01978365384616, "packet_end": 442.20776923076926, "raw_text": "通过高效的因果自注意力机制处理输入序列", "selected_text": "通过高效的因果自注意力机制处理输入序列", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 205 ], "packet_id": 15, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 434.03978365384614, "end": 436.25978365384617, "text": "生成提示词的上下文化表示", "raw_asr_text": "生成提示词的上下文化表示", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 15, "words": [ { "word": "生", "start": 434.03978365384614, "end": 434.35978365384614, "probability": 0.98779296875 }, { "word": "成", "start": 434.35978365384614, "end": 434.53978365384614, "probability": 0.99853515625 }, { "word": "提", "start": 434.53978365384614, "end": 434.69978365384617, "probability": 0.9990234375 }, { "word": "示", "start": 434.69978365384617, "end": 434.8797836538462, "probability": 1.0 }, { "word": "词", "start": 434.8797836538462, "end": 435.01978365384616, "probability": 0.998046875 }, { "word": "的", "start": 435.01978365384616, "end": 435.1797836538462, "probability": 0.99951171875 }, { "word": "上", "start": 435.1797836538462, "end": 435.3797836538462, "probability": 0.99951171875 }, { "word": "下", "start": 435.3797836538462, "end": 435.5597836538462, "probability": 1.0 }, { "word": "文", "start": 435.5597836538462, "end": 435.7397836538462, "probability": 1.0 }, { "word": "化", "start": 435.7397836538462, "end": 435.91978365384614, "probability": 0.9990234375 }, { "word": "表示", "start": 435.91978365384614, "end": 436.25978365384617, "probability": 0.97900390625 } ], "asr_metrics": { "avg_logprob": -0.05282413679669161, "compression_ratio": 1.3504043126684635, "no_speech_prob": 1.997897333683163e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "Yoko方面发现在常规Transformer架构中每一层都会生成并存储专属的全局KB缓存所以如果模型有40层你实际上要保存40份不同版本的记忆数据为了解决这种重复存储的问题他们提出了名为解码器解码器架构的解决方案模型的下半部分是自解码器通过高效的因果自注意力机制处理输入序列生成提示词的上下文化表示随后Yoko基于下半部分的输出生成共享的全局KB缓存模型上半部分则为交叉" } ], "id": 206, "review_annotation": "", "row_id": 206, "bounded_row_id": 206, "parent_row_ids": [ 206 ], "parent_row_id": 206, "packet_id": 15, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 413.01978365384616, "packet_end": 442.20776923076926, "raw_text": "生成提示词的上下文化表示", "selected_text": "生成提示词的上下文化表示", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 206 ], "packet_id": 15, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 436.25978365384617, "end": 440.33978365384615, "text": "随后Yoko基于下半部分的输出生成共享的全局KB缓存", "raw_asr_text": "随后Yoko基于下半部分的输出生成共享的全局KB缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 15, "words": [ { "word": "随", "start": 436.25978365384617, "end": 436.6797836538462, "probability": 0.998046875 }, { "word": "后", "start": 436.6797836538462, "end": 436.79978365384613, "probability": 0.9990234375 }, { "word": "Y", "start": 436.79978365384613, "end": 436.95978365384616, "probability": 0.9814453125 }, { "word": "oko", "start": 436.95978365384616, "end": 437.15978365384615, "probability": 0.99951171875 }, { "word": "基", "start": 437.15978365384615, "end": 437.35978365384614, "probability": 0.998046875 }, { "word": "于", "start": 437.35978365384614, "end": 437.47978365384614, "probability": 0.99951171875 }, { "word": "下", "start": 437.47978365384614, "end": 437.6797836538462, "probability": 1.0 }, { "word": "半", "start": 437.6797836538462, "end": 437.85978365384614, "probability": 1.0 }, { "word": "部分", "start": 437.85978365384614, "end": 438.09978365384615, "probability": 1.0 }, { "word": "的", "start": 438.09978365384615, "end": 438.19978365384617, "probability": 0.99365234375 }, { "word": "输", "start": 438.19978365384617, "end": 438.35978365384614, "probability": 0.99951171875 }, { "word": "出", "start": 438.35978365384614, "end": 438.4997836538462, "probability": 0.99951171875 }, { "word": "生", "start": 438.4997836538462, "end": 438.65978365384615, "probability": 0.99853515625 }, { "word": "成", "start": 438.65978365384615, "end": 438.83978365384615, "probability": 0.9990234375 }, { "word": "共", "start": 438.83978365384615, "end": 439.01978365384616, "probability": 0.99951171875 }, { "word": "享", "start": 439.01978365384616, "end": 439.13978365384617, "probability": 0.9912109375 }, { "word": "的", "start": 439.13978365384617, "end": 439.25978365384617, "probability": 0.99951171875 }, { "word": "全", "start": 439.25978365384617, "end": 439.51978365384616, "probability": 1.0 }, { "word": "局", "start": 439.51978365384616, "end": 439.6797836538462, "probability": 0.99951171875 }, { "word": "K", "start": 439.6797836538462, "end": 439.85978365384614, "probability": 0.99853515625 }, { "word": "B", "start": 439.85978365384614, "end": 439.95978365384616, "probability": 1.0 }, { "word": "缓", "start": 439.95978365384616, "end": 440.1197836538462, "probability": 1.0 }, { "word": "存", "start": 440.1197836538462, "end": 440.33978365384615, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.05282413679669161, "compression_ratio": 1.3504043126684635, "no_speech_prob": 1.997897333683163e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "Yoko方面发现在常规Transformer架构中每一层都会生成并存储专属的全局KB缓存所以如果模型有40层你实际上要保存40份不同版本的记忆数据为了解决这种重复存储的问题他们提出了名为解码器解码器架构的解决方案模型的下半部分是自解码器通过高效的因果自注意力机制处理输入序列生成提示词的上下文化表示随后Yoko基于下半部分的输出生成共享的全局KB缓存模型上半部分则为交叉" } ], "id": 207, "review_annotation": "", "row_id": 207, "bounded_row_id": 207, "parent_row_ids": [ 207 ], "parent_row_id": 207, "packet_id": 15, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 413.01978365384616, "packet_end": 442.20776923076926, "raw_text": "随后Yoko基于下半部分的输出生成共享的全局KB缓存", "selected_text": "随后Yoko基于下半部分的输出生成共享的全局KB缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 207 ], "packet_id": 15, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 440.33978365384615, "end": 441.40776923076925, "text": "模型上半部分则为交叉", "raw_asr_text": "模型上半部分则为交叉", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 15, "words": [ { "word": "模", "start": 440.33978365384615, "end": 440.59978365384615, "probability": 0.99951171875 }, { "word": "型", "start": 440.59978365384615, "end": 440.77978365384615, "probability": 1.0 }, { "word": "上", "start": 440.77978365384615, "end": 440.9397836538462, "probability": 0.99755859375 }, { "word": "半", "start": 440.9397836538462, "end": 441.0597836538462, "probability": 1.0 }, { "word": "部分", "start": 441.0597836538462, "end": 441.3797836538462, "probability": 1.0 }, { "word": "则", "start": 441.3797836538462, "end": 441.57978365384616, "probability": 0.99609375 }, { "word": "为", "start": 441.57978365384616, "end": 441.75978365384617, "probability": 0.99853515625 }, { "word": "交", "start": 441.75978365384617, "end": 441.9397836538462, "probability": 0.99853515625 }, { "word": "叉", "start": 441.9397836538462, "end": 442.1797836538462, "probability": 0.9912109375 } ], "asr_metrics": { "avg_logprob": -0.05282413679669161, "compression_ratio": 1.3504043126684635, "no_speech_prob": 1.997897333683163e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "Yoko方面发现在常规Transformer架构中每一层都会生成并存储专属的全局KB缓存所以如果模型有40层你实际上要保存40份不同版本的记忆数据为了解决这种重复存储的问题他们提出了名为解码器解码器架构的解决方案模型的下半部分是自解码器通过高效的因果自注意力机制处理输入序列生成提示词的上下文化表示随后Yoko基于下半部分的输出生成共享的全局KB缓存模型上半部分则为交叉" } ], "id": 208, "review_annotation": "", "row_id": 208, "bounded_row_id": 208, "parent_row_ids": [ 208 ], "parent_row_id": 208, "packet_id": 15, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 413.01978365384616, "packet_end": 442.20776923076926, "raw_text": "模型上半部分则为交叉", "selected_text": "模型上半部分则为交叉", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 208 ], "packet_id": 15, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 441.40776923076925, "end": 442.8077692307692, "text": "模型上半部分则为交叉解码器", "raw_asr_text": "模型上半部分则为交叉解码器", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "模", "start": 440.60776923076924, "end": 440.60776923076924, "probability": 0.3330078125 }, { "word": "型", "start": 440.60776923076924, "end": 440.76776923076926, "probability": 0.99609375 }, { "word": "上", "start": 440.76776923076926, "end": 440.9477692307692, "probability": 0.9853515625 }, { "word": "半", "start": 440.9477692307692, "end": 441.0677692307692, "probability": 0.9990234375 }, { "word": "部分", "start": 441.0677692307692, "end": 441.3677692307692, "probability": 0.99609375 }, { "word": "则", "start": 441.3677692307692, "end": 441.58776923076925, "probability": 0.978515625 }, { "word": "为", "start": 441.58776923076925, "end": 441.76776923076926, "probability": 0.97314453125 }, { "word": "交", "start": 441.76776923076926, "end": 441.92776923076923, "probability": 0.97216796875 }, { "word": "叉", "start": 441.92776923076923, "end": 442.26776923076926, "probability": 0.998046875 }, { "word": "解", "start": 442.26776923076926, "end": 442.4477692307692, "probability": 0.91015625 }, { "word": "码", "start": 442.4477692307692, "end": 442.64776923076926, "probability": 0.998046875 }, { "word": "器", "start": 442.64776923076926, "end": 442.8077692307692, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.1306228552924262, "compression_ratio": 1.345646437994723, "no_speech_prob": 5.959094329099912e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 209, "review_annotation": "", "row_id": 209, "bounded_row_id": 209, "parent_row_ids": [ 209 ], "parent_row_id": 209, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "模型上半部分则为交叉解码器", "selected_text": "模型上半部分则为交叉解码器", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 209 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 442.8077692307692, "end": 444.76776923076926, "text": "其中每一层都有独立的query", "raw_asr_text": "其中每一层都有独立的query", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "其", "start": 442.8077692307692, "end": 443.08776923076925, "probability": 0.96044921875 }, { "word": "中", "start": 443.08776923076925, "end": 443.2477692307692, "probability": 0.99951171875 }, { "word": "每", "start": 443.2477692307692, "end": 443.3877692307692, "probability": 0.9951171875 }, { "word": "一", "start": 443.3877692307692, "end": 443.5077692307692, "probability": 0.99072265625 }, { "word": "层", "start": 443.5077692307692, "end": 443.64776923076926, "probability": 0.9990234375 }, { "word": "都有", "start": 443.64776923076926, "end": 443.8677692307692, "probability": 0.9892578125 }, { "word": "独", "start": 443.8677692307692, "end": 444.0677692307692, "probability": 0.998046875 }, { "word": "立", "start": 444.0677692307692, "end": 444.1877692307692, "probability": 0.9990234375 }, { "word": "的", "start": 444.1877692307692, "end": 444.32776923076926, "probability": 0.99365234375 }, { "word": "qu", "start": 444.32776923076926, "end": 444.52776923076925, "probability": 0.6865234375 }, { "word": "ery", "start": 444.52776923076925, "end": 444.76776923076926, "probability": 0.9755859375 } ], "asr_metrics": { "avg_logprob": -0.1306228552924262, "compression_ratio": 1.345646437994723, "no_speech_prob": 5.959094329099912e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 210, "review_annotation": "", "row_id": 210, "bounded_row_id": 210, "parent_row_ids": [ 210 ], "parent_row_id": 210, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "其中每一层都有独立的query", "selected_text": "其中每一层都有独立的query", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 210 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 444.76776923076926, "end": 447.5677692307692, "text": "但所有层都会对同一份共享的k和v", "raw_asr_text": "但所有层都会对同一份共享的k和v", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "但", "start": 444.76776923076926, "end": 444.92776923076923, "probability": 0.91455078125 }, { "word": "所有", "start": 444.92776923076923, "end": 445.20776923076926, "probability": 0.99658203125 }, { "word": "层", "start": 445.20776923076926, "end": 445.4477692307692, "probability": 0.99853515625 }, { "word": "都", "start": 445.4477692307692, "end": 445.60776923076924, "probability": 0.9931640625 }, { "word": "会", "start": 445.60776923076924, "end": 445.76776923076926, "probability": 0.9990234375 }, { "word": "对", "start": 445.76776923076926, "end": 445.8877692307692, "probability": 0.99658203125 }, { "word": "同", "start": 445.8877692307692, "end": 446.1277692307692, "probability": 0.99072265625 }, { "word": "一", "start": 446.1277692307692, "end": 446.26776923076926, "probability": 0.998046875 }, { "word": "份", "start": 446.26776923076926, "end": 446.42776923076923, "probability": 0.9970703125 }, { "word": "共", "start": 446.42776923076923, "end": 446.6277692307692, "probability": 0.99560546875 }, { "word": "享", "start": 446.6277692307692, "end": 446.82776923076926, "probability": 0.99267578125 }, { "word": "的", "start": 446.82776923076926, "end": 446.96776923076925, "probability": 0.994140625 }, { "word": "k", "start": 446.96776923076925, "end": 447.20776923076926, "probability": 0.61083984375 }, { "word": "和", "start": 447.20776923076926, "end": 447.3877692307692, "probability": 0.875 }, { "word": "v", "start": 447.3877692307692, "end": 447.5677692307692, "probability": 0.9677734375 } ], "asr_metrics": { "avg_logprob": -0.1306228552924262, "compression_ratio": 1.345646437994723, "no_speech_prob": 5.959094329099912e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 211, "review_annotation": "", "row_id": 211, "bounded_row_id": 211, "parent_row_ids": [ 211 ], "parent_row_id": 211, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "但所有层都会对同一份共享的k和v", "selected_text": "但所有层都会对同一份共享的k和v", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 211 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 447.5677692307692, "end": 448.4477692307692, "text": "执行注意力计算", "raw_asr_text": "执行注意力计算", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "执", "start": 447.5677692307692, "end": 447.72776923076924, "probability": 0.79833984375 }, { "word": "行", "start": 447.72776923076924, "end": 447.8677692307692, "probability": 0.9990234375 }, { "word": "注意", "start": 447.8677692307692, "end": 448.04776923076923, "probability": 0.9853515625 }, { "word": "力", "start": 448.04776923076923, "end": 448.1877692307692, "probability": 0.998046875 }, { "word": "计", "start": 448.1877692307692, "end": 448.3077692307692, "probability": 0.998046875 }, { "word": "算", "start": 448.3077692307692, "end": 448.4477692307692, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1306228552924262, "compression_ratio": 1.345646437994723, "no_speech_prob": 5.959094329099912e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 212, "review_annotation": "", "row_id": 212, "bounded_row_id": 212, "parent_row_ids": [ 212 ], "parent_row_id": 212, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "执行注意力计算", "selected_text": "执行注意力计算", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 212 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 448.4477692307692, "end": 450.34776923076925, "text": "这样一来无需每一层都存储", "raw_asr_text": "这样一来无需每一层都存储", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "这样", "start": 448.4477692307692, "end": 448.78776923076924, "probability": 0.99072265625 }, { "word": "一", "start": 448.78776923076924, "end": 448.92776923076923, "probability": 0.92333984375 }, { "word": "来", "start": 448.92776923076923, "end": 449.0677692307692, "probability": 0.99951171875 }, { "word": "无", "start": 449.0677692307692, "end": 449.22776923076924, "probability": 0.96044921875 }, { "word": "需", "start": 449.22776923076924, "end": 449.3677692307692, "probability": 0.93994140625 }, { "word": "每", "start": 449.3677692307692, "end": 449.52776923076925, "probability": 0.99853515625 }, { "word": "一", "start": 449.52776923076925, "end": 449.66776923076924, "probability": 0.99365234375 }, { "word": "层", "start": 449.66776923076924, "end": 449.78776923076924, "probability": 1.0 }, { "word": "都", "start": 449.78776923076924, "end": 449.96776923076925, "probability": 0.99462890625 }, { "word": "存", "start": 449.96776923076925, "end": 450.16776923076924, "probability": 0.9990234375 }, { "word": "储", "start": 450.16776923076924, "end": 450.34776923076925, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.1306228552924262, "compression_ratio": 1.345646437994723, "no_speech_prob": 5.959094329099912e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 213, "review_annotation": "", "row_id": 213, "bounded_row_id": 213, "parent_row_ids": [ 213 ], "parent_row_id": 213, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "这样一来无需每一层都存储", "selected_text": "这样一来无需每一层都存储", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 213 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 450.34776923076925, "end": 452.1277692307692, "text": "各自体量庞大的kv缓存", "raw_asr_text": "各自体量庞大的kv缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "各", "start": 450.34776923076925, "end": 450.52776923076925, "probability": 0.94970703125 }, { "word": "自", "start": 450.52776923076925, "end": 450.70776923076926, "probability": 0.9951171875 }, { "word": "体", "start": 450.70776923076926, "end": 450.8077692307692, "probability": 0.99755859375 }, { "word": "量", "start": 450.8077692307692, "end": 450.96776923076925, "probability": 0.99658203125 }, { "word": "庞", "start": 450.96776923076925, "end": 451.20776923076926, "probability": 0.9736328125 }, { "word": "大的", "start": 451.20776923076926, "end": 451.4477692307692, "probability": 0.99853515625 }, { "word": "k", "start": 451.4477692307692, "end": 451.6277692307692, "probability": 0.9248046875 }, { "word": "v", "start": 451.6277692307692, "end": 451.7477692307692, "probability": 0.83544921875 }, { "word": "缓", "start": 451.7477692307692, "end": 451.90776923076925, "probability": 0.9951171875 }, { "word": "存", "start": 451.90776923076925, "end": 452.1277692307692, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1306228552924262, "compression_ratio": 1.345646437994723, "no_speech_prob": 5.959094329099912e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 214, "review_annotation": "", "row_id": 214, "bounded_row_id": 214, "parent_row_ids": [ 214 ], "parent_row_id": 214, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "各自体量庞大的kv缓存", "selected_text": "各自体量庞大的kv缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 214 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 452.1277692307692, "end": 454.3077692307692, "text": "Yoko基本上只需要生成一次全局记忆", "raw_asr_text": "Yoko基本上只需要生成一次全局记忆", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "Y", "start": 452.1277692307692, "end": 452.28776923076924, "probability": 0.4375 }, { "word": "oko", "start": 452.28776923076924, "end": 452.42776923076923, "probability": 0.8662109375 }, { "word": "基本", "start": 452.42776923076923, "end": 452.64776923076926, "probability": 0.9755859375 }, { "word": "上", "start": 452.64776923076926, "end": 452.8077692307692, "probability": 0.98828125 }, { "word": "只", "start": 452.8077692307692, "end": 452.9477692307692, "probability": 0.99755859375 }, { "word": "需要", "start": 452.9477692307692, "end": 453.1277692307692, "probability": 0.9873046875 }, { "word": "生", "start": 453.1277692307692, "end": 453.26776923076926, "probability": 0.9951171875 }, { "word": "成", "start": 453.26776923076926, "end": 453.40776923076925, "probability": 0.94775390625 }, { "word": "一次", "start": 453.40776923076925, "end": 453.70776923076926, "probability": 0.984375 }, { "word": "全", "start": 453.70776923076926, "end": 453.9477692307692, "probability": 0.8388671875 }, { "word": "局", "start": 453.9477692307692, "end": 454.04776923076923, "probability": 0.7001953125 }, { "word": "记", "start": 454.04776923076923, "end": 454.1877692307692, "probability": 0.736328125 }, { "word": "忆", "start": 454.1877692307692, "end": 454.3077692307692, "probability": 0.99267578125 } ], "asr_metrics": { "avg_logprob": -0.1306228552924262, "compression_ratio": 1.345646437994723, "no_speech_prob": 5.959094329099912e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 215, "review_annotation": "", "row_id": 215, "bounded_row_id": 215, "parent_row_ids": [ 215 ], "parent_row_id": 215, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "Yoko基本上只需要生成一次全局记忆", "selected_text": "Yoko基本上只需要生成一次全局记忆", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 215 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 454.3077692307692, "end": 455.52776923076925, "text": "就能供所有层服用", "raw_asr_text": "就能供所有层服用", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "就", "start": 454.3077692307692, "end": 454.48776923076923, "probability": 0.982421875 }, { "word": "能", "start": 454.48776923076923, "end": 454.6277692307692, "probability": 0.99951171875 }, { "word": "供", "start": 454.6277692307692, "end": 454.76776923076926, "probability": 0.93994140625 }, { "word": "所有", "start": 454.76776923076926, "end": 455.04776923076923, "probability": 0.9951171875 }, { "word": "层", "start": 455.04776923076923, "end": 455.22776923076924, "probability": 0.99951171875 }, { "word": "服", "start": 455.22776923076924, "end": 455.34776923076925, "probability": 0.37109375 }, { "word": "用", "start": 455.34776923076925, "end": 455.52776923076925, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1306228552924262, "compression_ratio": 1.345646437994723, "no_speech_prob": 5.959094329099912e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 216, "review_annotation": "", "row_id": 216, "bounded_row_id": 216, "parent_row_ids": [ 216 ], "parent_row_id": 216, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "就能供所有层服用", "selected_text": "就能供所有层服用", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 216 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 455.52776923076925, "end": 458.26776923076926, "text": "这也是其you only cache once命名的由来", "raw_asr_text": "这也是其you only cache once命名的由来", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "这", "start": 455.52776923076925, "end": 455.8877692307692, "probability": 0.98828125 }, { "word": "也是", "start": 455.8877692307692, "end": 456.1277692307692, "probability": 0.99609375 }, { "word": "其", "start": 456.1277692307692, "end": 456.34776923076925, "probability": 0.9931640625 }, { "word": "you", "start": 456.34776923076925, "end": 456.60776923076924, "probability": 0.2939453125 }, { "word": " only", "start": 456.60776923076924, "end": 456.8877692307692, "probability": 0.90869140625 }, { "word": " cache", "start": 456.8877692307692, "end": 457.1877692307692, "probability": 0.7724609375 }, { "word": " once", "start": 457.1877692307692, "end": 457.48776923076923, "probability": 0.83984375 }, { "word": "命", "start": 457.48776923076923, "end": 457.70776923076926, "probability": 0.9619140625 }, { "word": "名", "start": 457.70776923076926, "end": 457.82776923076926, "probability": 0.998046875 }, { "word": "的", "start": 457.82776923076926, "end": 457.92776923076923, "probability": 0.9970703125 }, { "word": "由", "start": 457.92776923076923, "end": 458.0677692307692, "probability": 0.9970703125 }, { "word": "来", "start": 458.0677692307692, "end": 458.26776923076926, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1306228552924262, "compression_ratio": 1.345646437994723, "no_speech_prob": 5.959094329099912e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 217, "review_annotation": "", "row_id": 217, "bounded_row_id": 217, "parent_row_ids": [ 217 ], "parent_row_id": 217, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "这也是其you only cache once命名的由来", "selected_text": "这也是其you only cache once命名的由来", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 217 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 458.26776923076926, "end": 460.16776923076924, "text": "尽管Yoko将其称为解码器", "raw_asr_text": "尽管Yoko将其称为解码器", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "尽", "start": 458.26776923076926, "end": 458.60776923076924, "probability": 0.990234375 }, { "word": "管", "start": 458.60776923076924, "end": 458.78776923076924, "probability": 0.99951171875 }, { "word": "Y", "start": 458.78776923076924, "end": 458.9477692307692, "probability": 0.92626953125 }, { "word": "oko", "start": 458.9477692307692, "end": 459.10776923076924, "probability": 0.998046875 }, { "word": "将", "start": 459.10776923076924, "end": 459.26776923076926, "probability": 0.9892578125 }, { "word": "其", "start": 459.26776923076926, "end": 459.40776923076925, "probability": 0.99560546875 }, { "word": "称", "start": 459.40776923076925, "end": 459.54776923076923, "probability": 0.9990234375 }, { "word": "为", "start": 459.54776923076923, "end": 459.6877692307692, "probability": 0.99169921875 }, { "word": "解", "start": 459.6877692307692, "end": 459.82776923076926, "probability": 0.98046875 }, { "word": "码", "start": 459.82776923076926, "end": 460.0077692307692, "probability": 1.0 }, { "word": "器", "start": 460.0077692307692, "end": 460.16776923076924, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1306228552924262, "compression_ratio": 1.345646437994723, "no_speech_prob": 5.959094329099912e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 218, "review_annotation": "", "row_id": 218, "bounded_row_id": 218, "parent_row_ids": [ 218 ], "parent_row_id": 218, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "尽管Yoko将其称为解码器", "selected_text": "尽管Yoko将其称为解码器", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 218 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 460.16776923076924, "end": 461.08776923076925, "text": "解码器架构", "raw_asr_text": "解码器架构", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "解", "start": 460.16776923076924, "end": 460.3677692307692, "probability": 0.61279296875 }, { "word": "码", "start": 460.3677692307692, "end": 460.54776923076923, "probability": 1.0 }, { "word": "器", "start": 460.54776923076923, "end": 460.6877692307692, "probability": 0.99951171875 }, { "word": "架", "start": 460.6877692307692, "end": 460.90776923076925, "probability": 0.9990234375 }, { "word": "构", "start": 460.90776923076925, "end": 461.08776923076925, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.1306228552924262, "compression_ratio": 1.345646437994723, "no_speech_prob": 5.959094329099912e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 219, "review_annotation": "", "row_id": 219, "bounded_row_id": 219, "parent_row_ids": [ 219 ], "parent_row_id": 219, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "解码器架构", "selected_text": "解码器架构", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 219 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 461.08776923076925, "end": 462.58776923076925, "text": "但它的训练和使用方式", "raw_asr_text": "但它的训练和使用方式", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "但", "start": 461.08776923076925, "end": 461.28776923076924, "probability": 0.9853515625 }, { "word": "它的", "start": 461.28776923076924, "end": 461.48776923076923, "probability": 0.8603515625 }, { "word": "训", "start": 461.48776923076923, "end": 461.66776923076924, "probability": 0.9970703125 }, { "word": "练", "start": 461.66776923076924, "end": 461.8077692307692, "probability": 1.0 }, { "word": "和", "start": 461.8077692307692, "end": 461.92776923076923, "probability": 0.99609375 }, { "word": "使", "start": 461.92776923076923, "end": 462.08776923076925, "probability": 0.9970703125 }, { "word": "用", "start": 462.08776923076925, "end": 462.20776923076926, "probability": 1.0 }, { "word": "方", "start": 462.20776923076926, "end": 462.3877692307692, "probability": 0.9990234375 }, { "word": "式", "start": 462.3877692307692, "end": 462.58776923076925, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.1306228552924262, "compression_ratio": 1.345646437994723, "no_speech_prob": 5.959094329099912e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 220, "review_annotation": "", "row_id": 220, "bounded_row_id": 220, "parent_row_ids": [ 220 ], "parent_row_id": 220, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "但它的训练和使用方式", "selected_text": "但它的训练和使用方式", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 220 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 462.58776923076925, "end": 464.76776923076926, "text": "仍与常规自回规语言模型一致", "raw_asr_text": "仍与常规自回规语言模型一致", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "仍", "start": 462.58776923076925, "end": 462.82776923076926, "probability": 0.99658203125 }, { "word": "与", "start": 462.82776923076926, "end": 462.98776923076923, "probability": 0.9677734375 }, { "word": "常", "start": 462.98776923076923, "end": 463.16776923076924, "probability": 0.9970703125 }, { "word": "规", "start": 463.16776923076924, "end": 463.40776923076925, "probability": 0.99951171875 }, { "word": "自", "start": 463.40776923076925, "end": 463.5677692307692, "probability": 0.8681640625 }, { "word": "回", "start": 463.5677692307692, "end": 463.6877692307692, "probability": 0.45263671875 }, { "word": "规", "start": 463.6877692307692, "end": 463.82776923076926, "probability": 0.80126953125 }, { "word": "语", "start": 463.82776923076926, "end": 463.98776923076923, "probability": 0.9677734375 }, { "word": "言", "start": 463.98776923076923, "end": 464.08776923076925, "probability": 0.99755859375 }, { "word": "模", "start": 464.08776923076925, "end": 464.22776923076924, "probability": 0.9970703125 }, { "word": "型", "start": 464.22776923076924, "end": 464.3877692307692, "probability": 1.0 }, { "word": "一", "start": 464.3877692307692, "end": 464.5677692307692, "probability": 0.998046875 }, { "word": "致", "start": 464.5677692307692, "end": 464.76776923076926, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1306228552924262, "compression_ratio": 1.345646437994723, "no_speech_prob": 5.959094329099912e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 221, "review_annotation": "", "row_id": 221, "bounded_row_id": 221, "parent_row_ids": [ 221 ], "parent_row_id": 221, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "仍与常规自回规语言模型一致", "selected_text": "仍与常规自回规语言模型一致", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 221 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 464.76776923076926, "end": 466.72776923076924, "text": "这个名称听起来可能和编码器", "raw_asr_text": "这个名称听起来可能和编码器", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "这个", "start": 464.76776923076926, "end": 465.08776923076925, "probability": 0.98974609375 }, { "word": "名", "start": 465.08776923076925, "end": 465.22776923076924, "probability": 0.998046875 }, { "word": "称", "start": 465.22776923076924, "end": 465.3877692307692, "probability": 0.9990234375 }, { "word": "听", "start": 465.3877692307692, "end": 465.52776923076925, "probability": 0.99755859375 }, { "word": "起", "start": 465.52776923076925, "end": 465.66776923076924, "probability": 0.998046875 }, { "word": "来", "start": 465.66776923076924, "end": 465.82776923076926, "probability": 1.0 }, { "word": "可能", "start": 465.82776923076926, "end": 466.04776923076923, "probability": 0.99462890625 }, { "word": "和", "start": 466.04776923076923, "end": 466.22776923076924, "probability": 0.99755859375 }, { "word": "编", "start": 466.22776923076924, "end": 466.3677692307692, "probability": 0.8671875 }, { "word": "码", "start": 466.3677692307692, "end": 466.5677692307692, "probability": 1.0 }, { "word": "器", "start": 466.5677692307692, "end": 466.72776923076924, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1306228552924262, "compression_ratio": 1.345646437994723, "no_speech_prob": 5.959094329099912e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 222, "review_annotation": "", "row_id": 222, "bounded_row_id": 222, "parent_row_ids": [ 222 ], "parent_row_id": 222, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "这个名称听起来可能和编码器", "selected_text": "这个名称听起来可能和编码器", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 222 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 466.72776923076924, "end": 468.04776923076923, "text": "解码器架构相似", "raw_asr_text": "解码器架构相似", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "解", "start": 466.72776923076924, "end": 466.96776923076925, "probability": 0.9921875 }, { "word": "码", "start": 466.96776923076925, "end": 467.1277692307692, "probability": 1.0 }, { "word": "器", "start": 467.1277692307692, "end": 467.2477692307692, "probability": 1.0 }, { "word": "架", "start": 467.2477692307692, "end": 467.46776923076925, "probability": 0.9990234375 }, { "word": "构", "start": 467.46776923076925, "end": 467.58776923076925, "probability": 1.0 }, { "word": "相", "start": 467.58776923076925, "end": 467.70776923076926, "probability": 0.98388671875 }, { "word": "似", "start": 467.70776923076926, "end": 468.04776923076923, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.1306228552924262, "compression_ratio": 1.345646437994723, "no_speech_prob": 5.959094329099912e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 223, "review_annotation": "", "row_id": 223, "bounded_row_id": 223, "parent_row_ids": [ 223 ], "parent_row_id": 223, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "解码器架构相似", "selected_text": "解码器架构相似", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 223 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 468.04776923076923, "end": 468.9957548076924, "text": "容易让人联想到他与最终的", "raw_asr_text": "容易让人联想到他与最终的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 16, "words": [ { "word": "容易", "start": 468.04776923076923, "end": 468.42776923076923, "probability": 0.99462890625 }, { "word": "让", "start": 468.42776923076923, "end": 468.6277692307692, "probability": 0.99462890625 }, { "word": "人", "start": 468.6277692307692, "end": 468.8077692307692, "probability": 0.99951171875 }, { "word": "联", "start": 468.8077692307692, "end": 469.0077692307692, "probability": 0.998046875 }, { "word": "想", "start": 469.0077692307692, "end": 469.1277692307692, "probability": 0.9990234375 }, { "word": "到", "start": 469.1277692307692, "end": 469.2477692307692, "probability": 0.9970703125 }, { "word": "他", "start": 469.2477692307692, "end": 469.40776923076925, "probability": 0.60205078125 }, { "word": "与", "start": 469.40776923076925, "end": 469.6277692307692, "probability": 0.990234375 }, { "word": "最", "start": 469.6277692307692, "end": 469.76776923076926, "probability": 0.84619140625 }, { "word": "终", "start": 469.76776923076926, "end": 469.76776923076926, "probability": 0.54150390625 }, { "word": "的", "start": 469.76776923076926, "end": 469.76776923076926, "probability": 0.54052734375 } ], "asr_metrics": { "avg_logprob": -0.3648134119370404, "compression_ratio": 0.7659574468085106, "no_speech_prob": 1.2412390559823905e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "模型上半部分则为交叉解码器其中每一层都有独立的query但所有层都会对同一份共享的k和v执行注意力计算这样一来无需每一层都存储各自体量庞大的kv缓存Yoko基本上只需要生成一次全局记忆就能供所有层服用这也是其you only cache once命名的由来尽管Yoko将其称为解码器解码器架构但它的训练和使用方式仍与常规自回规语言模型一致这个名称听起来可能和编码器解码器架构相似容易让人联想到他与最终的" } ], "id": 224, "review_annotation": "", "row_id": 224, "bounded_row_id": 224, "parent_row_ids": [ 224 ], "parent_row_id": 224, "packet_id": 16, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 440.60776923076924, "packet_end": 469.7957548076924, "raw_text": "容易让人联想到他与最终的", "selected_text": "容易让人联想到他与最终的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 224 ], "packet_id": 16, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 468.9957548076924, "end": 471.59575480769234, "text": "容易让人联想到它与最初的注意力机制存在关联", "raw_asr_text": "容易让人联想到它与最初的注意力机制存在关联", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 17, "words": [ { "word": "容易", "start": 468.19575480769237, "end": 468.39575480769236, "probability": 0.794921875 }, { "word": "让", "start": 468.39575480769236, "end": 468.63575480769236, "probability": 0.99072265625 }, { "word": "人", "start": 468.63575480769236, "end": 468.8157548076924, "probability": 0.99755859375 }, { "word": "联", "start": 468.8157548076924, "end": 468.9957548076924, "probability": 0.9970703125 }, { "word": "想", "start": 468.9957548076924, "end": 469.1157548076924, "probability": 0.9990234375 }, { "word": "到", "start": 469.1157548076924, "end": 469.2357548076924, "probability": 0.99267578125 }, { "word": "它", "start": 469.2357548076924, "end": 469.4157548076924, "probability": 0.912109375 }, { "word": "与", "start": 469.4157548076924, "end": 469.6157548076924, "probability": 0.9921875 }, { "word": "最", "start": 469.6157548076924, "end": 469.7357548076924, "probability": 0.99365234375 }, { "word": "初", "start": 469.7357548076924, "end": 469.95575480769236, "probability": 0.99853515625 }, { "word": "的", "start": 469.95575480769236, "end": 470.07575480769236, "probability": 0.99267578125 }, { "word": "注意", "start": 470.07575480769236, "end": 470.27575480769235, "probability": 0.98876953125 }, { "word": "力", "start": 470.27575480769235, "end": 470.47575480769234, "probability": 0.9990234375 }, { "word": "机", "start": 470.47575480769234, "end": 470.6157548076924, "probability": 0.994140625 }, { "word": "制", "start": 470.6157548076924, "end": 470.7957548076924, "probability": 0.98974609375 }, { "word": "存", "start": 470.7957548076924, "end": 471.0557548076924, "probability": 0.98779296875 }, { "word": "在", "start": 471.0557548076924, "end": 471.2357548076924, "probability": 0.99951171875 }, { "word": "关", "start": 471.2357548076924, "end": 471.39575480769236, "probability": 0.99462890625 }, { "word": "联", "start": 471.39575480769236, "end": 471.59575480769234, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.1252976717523566, "compression_ratio": 1.4666666666666666, "no_speech_prob": 2.5849721668746994e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "容易让人联想到它与最初的注意力机制存在关联但它取这个名字仅仅是因为架构包含两种不同类型的解码器下半部分名为自解码器是因为它采用因果资注意力处理序列构建共享的全局KV上半部分名为交叉解码器是因为每个上层都会生成独立的query并对下半部分生成的KV执行交叉注意力计算这种设计还有个更棒的好处既然上半部分只需要用到下半部分生成的KV那么在玉田充阶段严格来说" } ], "id": 225, "review_annotation": "", "row_id": 225, "bounded_row_id": 225, "parent_row_ids": [ 225 ], "parent_row_id": 225, "packet_id": 17, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 468.19575480769237, "packet_end": 497.3837403846154, "raw_text": "容易让人联想到它与最初的注意力机制存在关联", "selected_text": "容易让人联想到它与最初的注意力机制存在关联", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 225 ], "packet_id": 17, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 471.59575480769234, "end": 475.8757548076924, "text": "但它取这个名字仅仅是因为架构包含两种不同类型的解码器", "raw_asr_text": "但它取这个名字仅仅是因为架构包含两种不同类型的解码器", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 17, "words": [ { "word": "但", "start": 471.59575480769234, "end": 471.9357548076924, "probability": 0.9052734375 }, { "word": "它", "start": 471.9357548076924, "end": 472.1157548076924, "probability": 0.912109375 }, { "word": "取", "start": 472.1157548076924, "end": 472.33575480769235, "probability": 0.9921875 }, { "word": "这个", "start": 472.33575480769235, "end": 472.53575480769234, "probability": 0.95703125 }, { "word": "名", "start": 472.53575480769234, "end": 472.69575480769237, "probability": 0.99853515625 }, { "word": "字", "start": 472.69575480769237, "end": 472.8557548076924, "probability": 0.9970703125 }, { "word": "仅", "start": 472.8557548076924, "end": 473.0557548076924, "probability": 0.9931640625 }, { "word": "仅", "start": 473.0557548076924, "end": 473.13575480769236, "probability": 0.9990234375 }, { "word": "是", "start": 473.13575480769236, "end": 473.3157548076924, "probability": 0.98974609375 }, { "word": "因为", "start": 473.3157548076924, "end": 473.51575480769236, "probability": 0.986328125 }, { "word": "架", "start": 473.51575480769236, "end": 473.75575480769237, "probability": 0.998046875 }, { "word": "构", "start": 473.75575480769237, "end": 473.9357548076924, "probability": 0.99853515625 }, { "word": "包", "start": 473.9357548076924, "end": 474.1157548076924, "probability": 0.9912109375 }, { "word": "含", "start": 474.1157548076924, "end": 474.27575480769235, "probability": 0.9990234375 }, { "word": "两", "start": 474.27575480769235, "end": 474.45575480769236, "probability": 0.98388671875 }, { "word": "种", "start": 474.45575480769236, "end": 474.69575480769237, "probability": 0.9990234375 }, { "word": "不同", "start": 474.69575480769237, "end": 474.97575480769234, "probability": 0.9990234375 }, { "word": "类", "start": 474.97575480769234, "end": 475.2357548076924, "probability": 1.0 }, { "word": "型", "start": 475.2357548076924, "end": 475.3757548076924, "probability": 0.99853515625 }, { "word": "的", "start": 475.3757548076924, "end": 475.47575480769234, "probability": 0.98388671875 }, { "word": "解", "start": 475.47575480769234, "end": 475.57575480769236, "probability": 0.9912109375 }, { "word": "码", "start": 475.57575480769236, "end": 475.75575480769237, "probability": 0.9970703125 }, { "word": "器", "start": 475.75575480769237, "end": 475.8757548076924, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1252976717523566, "compression_ratio": 1.4666666666666666, "no_speech_prob": 2.5849721668746994e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "容易让人联想到它与最初的注意力机制存在关联但它取这个名字仅仅是因为架构包含两种不同类型的解码器下半部分名为自解码器是因为它采用因果资注意力处理序列构建共享的全局KV上半部分名为交叉解码器是因为每个上层都会生成独立的query并对下半部分生成的KV执行交叉注意力计算这种设计还有个更棒的好处既然上半部分只需要用到下半部分生成的KV那么在玉田充阶段严格来说" } ], "id": 226, "review_annotation": "", "row_id": 226, "bounded_row_id": 226, "parent_row_ids": [ 226 ], "parent_row_id": 226, "packet_id": 17, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 468.19575480769237, "packet_end": 497.3837403846154, "raw_text": "但它取这个名字仅仅是因为架构包含两种不同类型的解码器", "selected_text": "但它取这个名字仅仅是因为架构包含两种不同类型的解码器", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 226 ], "packet_id": 17, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 475.8757548076924, "end": 477.75575480769237, "text": "下半部分名为自解码器", "raw_asr_text": "下半部分名为自解码器", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 17, "words": [ { "word": "下", "start": 475.8757548076924, "end": 476.13575480769236, "probability": 0.98583984375 }, { "word": "半", "start": 476.13575480769236, "end": 476.27575480769235, "probability": 1.0 }, { "word": "部分", "start": 476.27575480769235, "end": 476.5557548076924, "probability": 0.9951171875 }, { "word": "名", "start": 476.5557548076924, "end": 476.77575480769235, "probability": 0.98828125 }, { "word": "为", "start": 476.77575480769235, "end": 477.01575480769236, "probability": 0.998046875 }, { "word": "自", "start": 477.01575480769236, "end": 477.21575480769235, "probability": 0.76708984375 }, { "word": "解", "start": 477.21575480769235, "end": 477.3757548076924, "probability": 0.99462890625 }, { "word": "码", "start": 477.3757548076924, "end": 477.57575480769236, "probability": 1.0 }, { "word": "器", "start": 477.57575480769236, "end": 477.75575480769237, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1252976717523566, "compression_ratio": 1.4666666666666666, "no_speech_prob": 2.5849721668746994e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "容易让人联想到它与最初的注意力机制存在关联但它取这个名字仅仅是因为架构包含两种不同类型的解码器下半部分名为自解码器是因为它采用因果资注意力处理序列构建共享的全局KV上半部分名为交叉解码器是因为每个上层都会生成独立的query并对下半部分生成的KV执行交叉注意力计算这种设计还有个更棒的好处既然上半部分只需要用到下半部分生成的KV那么在玉田充阶段严格来说" } ], "id": 227, "review_annotation": "", "row_id": 227, "bounded_row_id": 227, "parent_row_ids": [ 227 ], "parent_row_id": 227, "packet_id": 17, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 468.19575480769237, "packet_end": 497.3837403846154, "raw_text": "下半部分名为自解码器", "selected_text": "下半部分名为自解码器", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 227 ], "packet_id": 17, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 477.75575480769237, "end": 480.4957548076924, "text": "是因为它采用因果资注意力处理序列", "raw_asr_text": "是因为它采用因果资注意力处理序列", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 17, "words": [ { "word": "是", "start": 477.75575480769237, "end": 478.03575480769234, "probability": 0.98046875 }, { "word": "因为", "start": 478.03575480769234, "end": 478.27575480769235, "probability": 0.99365234375 }, { "word": "它", "start": 478.27575480769235, "end": 478.47575480769234, "probability": 0.982421875 }, { "word": "采", "start": 478.47575480769234, "end": 478.6757548076924, "probability": 0.998046875 }, { "word": "用", "start": 478.6757548076924, "end": 478.83575480769235, "probability": 0.99951171875 }, { "word": "因", "start": 478.83575480769235, "end": 479.03575480769234, "probability": 0.98876953125 }, { "word": "果", "start": 479.03575480769234, "end": 479.2357548076924, "probability": 1.0 }, { "word": "资", "start": 479.2357548076924, "end": 479.4157548076924, "probability": 0.748046875 }, { "word": "注意", "start": 479.4157548076924, "end": 479.63575480769236, "probability": 0.9287109375 }, { "word": "力", "start": 479.63575480769236, "end": 479.8157548076924, "probability": 0.99951171875 }, { "word": "处", "start": 479.8157548076924, "end": 480.01575480769236, "probability": 0.99755859375 }, { "word": "理", "start": 480.01575480769236, "end": 480.15575480769235, "probability": 1.0 }, { "word": "序", "start": 480.15575480769235, "end": 480.3157548076924, "probability": 0.998046875 }, { "word": "列", "start": 480.3157548076924, "end": 480.4957548076924, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.1252976717523566, "compression_ratio": 1.4666666666666666, "no_speech_prob": 2.5849721668746994e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "容易让人联想到它与最初的注意力机制存在关联但它取这个名字仅仅是因为架构包含两种不同类型的解码器下半部分名为自解码器是因为它采用因果资注意力处理序列构建共享的全局KV上半部分名为交叉解码器是因为每个上层都会生成独立的query并对下半部分生成的KV执行交叉注意力计算这种设计还有个更棒的好处既然上半部分只需要用到下半部分生成的KV那么在玉田充阶段严格来说" } ], "id": 228, "review_annotation": "", "row_id": 228, "bounded_row_id": 228, "parent_row_ids": [ 228 ], "parent_row_id": 228, "packet_id": 17, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 468.19575480769237, "packet_end": 497.3837403846154, "raw_text": "是因为它采用因果资注意力处理序列", "selected_text": "是因为它采用因果资注意力处理序列", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 228 ], "packet_id": 17, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 480.4957548076924, "end": 482.3757548076924, "text": "构建共享的全局KV", "raw_asr_text": "构建共享的全局KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 17, "words": [ { "word": "构", "start": 480.4957548076924, "end": 480.7957548076924, "probability": 0.9853515625 }, { "word": "建", "start": 480.7957548076924, "end": 480.9957548076924, "probability": 0.99658203125 }, { "word": "共", "start": 480.9957548076924, "end": 481.27575480769235, "probability": 0.99267578125 }, { "word": "享", "start": 481.27575480769235, "end": 481.4357548076924, "probability": 0.99169921875 }, { "word": "的", "start": 481.4357548076924, "end": 481.59575480769234, "probability": 0.99560546875 }, { "word": "全", "start": 481.59575480769234, "end": 481.8157548076924, "probability": 0.99072265625 }, { "word": "局", "start": 481.8157548076924, "end": 481.9957548076924, "probability": 0.99853515625 }, { "word": "K", "start": 481.9957548076924, "end": 482.19575480769237, "probability": 0.80712890625 }, { "word": "V", "start": 482.19575480769237, "end": 482.3757548076924, "probability": 0.89306640625 } ], "asr_metrics": { "avg_logprob": -0.1252976717523566, "compression_ratio": 1.4666666666666666, "no_speech_prob": 2.5849721668746994e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "容易让人联想到它与最初的注意力机制存在关联但它取这个名字仅仅是因为架构包含两种不同类型的解码器下半部分名为自解码器是因为它采用因果资注意力处理序列构建共享的全局KV上半部分名为交叉解码器是因为每个上层都会生成独立的query并对下半部分生成的KV执行交叉注意力计算这种设计还有个更棒的好处既然上半部分只需要用到下半部分生成的KV那么在玉田充阶段严格来说" } ], "id": 229, "review_annotation": "", "row_id": 229, "bounded_row_id": 229, "parent_row_ids": [ 229 ], "parent_row_id": 229, "packet_id": 17, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 468.19575480769237, "packet_end": 497.3837403846154, "raw_text": "构建共享的全局KV", "selected_text": "构建共享的全局KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 229 ], "packet_id": 17, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 482.3757548076924, "end": 484.51575480769236, "text": "上半部分名为交叉解码器", "raw_asr_text": "上半部分名为交叉解码器", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 17, "words": [ { "word": "上", "start": 482.3757548076924, "end": 482.7957548076924, "probability": 0.8779296875 }, { "word": "半", "start": 482.7957548076924, "end": 482.95575480769236, "probability": 1.0 }, { "word": "部分", "start": 482.95575480769236, "end": 483.21575480769235, "probability": 0.9990234375 }, { "word": "名", "start": 483.21575480769235, "end": 483.3557548076924, "probability": 0.9970703125 }, { "word": "为", "start": 483.3557548076924, "end": 483.53575480769234, "probability": 0.9990234375 }, { "word": "交", "start": 483.53575480769234, "end": 483.75575480769237, "probability": 0.998046875 }, { "word": "叉", "start": 483.75575480769237, "end": 483.97575480769234, "probability": 0.9970703125 }, { "word": "解", "start": 483.97575480769234, "end": 484.15575480769235, "probability": 0.99951171875 }, { "word": "码", "start": 484.15575480769235, "end": 484.33575480769235, "probability": 1.0 }, { "word": "器", "start": 484.33575480769235, "end": 484.51575480769236, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1252976717523566, "compression_ratio": 1.4666666666666666, "no_speech_prob": 2.5849721668746994e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "容易让人联想到它与最初的注意力机制存在关联但它取这个名字仅仅是因为架构包含两种不同类型的解码器下半部分名为自解码器是因为它采用因果资注意力处理序列构建共享的全局KV上半部分名为交叉解码器是因为每个上层都会生成独立的query并对下半部分生成的KV执行交叉注意力计算这种设计还有个更棒的好处既然上半部分只需要用到下半部分生成的KV那么在玉田充阶段严格来说" } ], "id": 230, "review_annotation": "", "row_id": 230, "bounded_row_id": 230, "parent_row_ids": [ 230 ], "parent_row_id": 230, "packet_id": 17, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 468.19575480769237, "packet_end": 497.3837403846154, "raw_text": "上半部分名为交叉解码器", "selected_text": "上半部分名为交叉解码器", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 230 ], "packet_id": 17, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 484.51575480769236, "end": 487.09575480769234, "text": "是因为每个上层都会生成独立的query", "raw_asr_text": "是因为每个上层都会生成独立的query", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 17, "words": [ { "word": "是", "start": 484.51575480769236, "end": 484.75575480769237, "probability": 0.99462890625 }, { "word": "因为", "start": 484.75575480769237, "end": 484.9357548076924, "probability": 0.99365234375 }, { "word": "每", "start": 484.9357548076924, "end": 485.15575480769235, "probability": 0.998046875 }, { "word": "个", "start": 485.15575480769235, "end": 485.3157548076924, "probability": 0.99755859375 }, { "word": "上", "start": 485.3157548076924, "end": 485.4957548076924, "probability": 0.99853515625 }, { "word": "层", "start": 485.4957548076924, "end": 485.65575480769235, "probability": 1.0 }, { "word": "都", "start": 485.65575480769235, "end": 485.8157548076924, "probability": 0.99853515625 }, { "word": "会", "start": 485.8157548076924, "end": 485.95575480769236, "probability": 0.99951171875 }, { "word": "生", "start": 485.95575480769236, "end": 486.07575480769236, "probability": 0.98291015625 }, { "word": "成", "start": 486.07575480769236, "end": 486.27575480769235, "probability": 0.99755859375 }, { "word": "独", "start": 486.27575480769235, "end": 486.53575480769234, "probability": 0.99951171875 }, { "word": "立", "start": 486.53575480769234, "end": 486.69575480769237, "probability": 0.99951171875 }, { "word": "的", "start": 486.69575480769237, "end": 486.83575480769235, "probability": 0.99853515625 }, { "word": "qu", "start": 486.83575480769235, "end": 486.91575480769234, "probability": 0.43359375 }, { "word": "ery", "start": 486.91575480769234, "end": 487.09575480769234, "probability": 0.9921875 } ], "asr_metrics": { "avg_logprob": -0.1252976717523566, "compression_ratio": 1.4666666666666666, "no_speech_prob": 2.5849721668746994e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "容易让人联想到它与最初的注意力机制存在关联但它取这个名字仅仅是因为架构包含两种不同类型的解码器下半部分名为自解码器是因为它采用因果资注意力处理序列构建共享的全局KV上半部分名为交叉解码器是因为每个上层都会生成独立的query并对下半部分生成的KV执行交叉注意力计算这种设计还有个更棒的好处既然上半部分只需要用到下半部分生成的KV那么在玉田充阶段严格来说" } ], "id": 231, "review_annotation": "", "row_id": 231, "bounded_row_id": 231, "parent_row_ids": [ 231 ], "parent_row_id": 231, "packet_id": 17, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 468.19575480769237, "packet_end": 497.3837403846154, "raw_text": "是因为每个上层都会生成独立的query", "selected_text": "是因为每个上层都会生成独立的query", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 231 ], "packet_id": 17, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 487.09575480769234, "end": 490.07575480769236, "text": "并对下半部分生成的KV执行交叉注意力计算", "raw_asr_text": "并对下半部分生成的KV执行交叉注意力计算", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 17, "words": [ { "word": "并", "start": 487.09575480769234, "end": 487.3557548076924, "probability": 0.9951171875 }, { "word": "对", "start": 487.3557548076924, "end": 487.4757548076924, "probability": 0.9990234375 }, { "word": "下", "start": 487.4757548076924, "end": 487.6757548076924, "probability": 0.99853515625 }, { "word": "半", "start": 487.6757548076924, "end": 487.7957548076924, "probability": 1.0 }, { "word": "部分", "start": 487.7957548076924, "end": 488.03575480769234, "probability": 0.99853515625 }, { "word": "生", "start": 488.03575480769234, "end": 488.19575480769237, "probability": 0.9951171875 }, { "word": "成", "start": 488.19575480769237, "end": 488.33575480769235, "probability": 0.994140625 }, { "word": "的", "start": 488.33575480769235, "end": 488.4357548076924, "probability": 0.99951171875 }, { "word": "K", "start": 488.4357548076924, "end": 488.57575480769236, "probability": 0.96533203125 }, { "word": "V", "start": 488.57575480769236, "end": 488.75575480769237, "probability": 0.9921875 }, { "word": "执", "start": 488.75575480769237, "end": 488.95575480769236, "probability": 0.939453125 }, { "word": "行", "start": 488.95575480769236, "end": 489.09575480769234, "probability": 0.9990234375 }, { "word": "交", "start": 489.09575480769234, "end": 489.25575480769237, "probability": 0.9990234375 }, { "word": "叉", "start": 489.25575480769237, "end": 489.4357548076924, "probability": 1.0 }, { "word": "注意", "start": 489.4357548076924, "end": 489.65575480769235, "probability": 0.99560546875 }, { "word": "力", "start": 489.65575480769235, "end": 489.77575480769235, "probability": 0.9990234375 }, { "word": "计", "start": 489.77575480769235, "end": 489.9357548076924, "probability": 0.99853515625 }, { "word": "算", "start": 489.9357548076924, "end": 490.07575480769236, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1252976717523566, "compression_ratio": 1.4666666666666666, "no_speech_prob": 2.5849721668746994e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "容易让人联想到它与最初的注意力机制存在关联但它取这个名字仅仅是因为架构包含两种不同类型的解码器下半部分名为自解码器是因为它采用因果资注意力处理序列构建共享的全局KV上半部分名为交叉解码器是因为每个上层都会生成独立的query并对下半部分生成的KV执行交叉注意力计算这种设计还有个更棒的好处既然上半部分只需要用到下半部分生成的KV那么在玉田充阶段严格来说" } ], "id": 232, "review_annotation": "", "row_id": 232, "bounded_row_id": 232, "parent_row_ids": [ 232 ], "parent_row_id": 232, "packet_id": 17, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 468.19575480769237, "packet_end": 497.3837403846154, "raw_text": "并对下半部分生成的KV执行交叉注意力计算", "selected_text": "并对下半部分生成的KV执行交叉注意力计算", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 232 ], "packet_id": 17, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 490.07575480769236, "end": 492.03575480769234, "text": "这种设计还有个更棒的好处", "raw_asr_text": "这种设计还有个更棒的好处", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 17, "words": [ { "word": "这", "start": 490.07575480769236, "end": 490.39575480769236, "probability": 0.9912109375 }, { "word": "种", "start": 490.39575480769236, "end": 490.51575480769236, "probability": 0.99853515625 }, { "word": "设", "start": 490.51575480769236, "end": 490.69575480769237, "probability": 1.0 }, { "word": "计", "start": 490.69575480769237, "end": 490.83575480769235, "probability": 0.99951171875 }, { "word": "还有", "start": 490.83575480769235, "end": 491.03575480769234, "probability": 0.99755859375 }, { "word": "个", "start": 491.03575480769234, "end": 491.2357548076924, "probability": 0.84423828125 }, { "word": "更", "start": 491.2357548076924, "end": 491.3757548076924, "probability": 0.998046875 }, { "word": "棒", "start": 491.3757548076924, "end": 491.5557548076924, "probability": 0.994140625 }, { "word": "的", "start": 491.5557548076924, "end": 491.65575480769235, "probability": 1.0 }, { "word": "好", "start": 491.65575480769235, "end": 491.83575480769235, "probability": 0.99951171875 }, { "word": "处", "start": 491.83575480769235, "end": 492.03575480769234, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1252976717523566, "compression_ratio": 1.4666666666666666, "no_speech_prob": 2.5849721668746994e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "容易让人联想到它与最初的注意力机制存在关联但它取这个名字仅仅是因为架构包含两种不同类型的解码器下半部分名为自解码器是因为它采用因果资注意力处理序列构建共享的全局KV上半部分名为交叉解码器是因为每个上层都会生成独立的query并对下半部分生成的KV执行交叉注意力计算这种设计还有个更棒的好处既然上半部分只需要用到下半部分生成的KV那么在玉田充阶段严格来说" } ], "id": 233, "review_annotation": "", "row_id": 233, "bounded_row_id": 233, "parent_row_ids": [ 233 ], "parent_row_id": 233, "packet_id": 17, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 468.19575480769237, "packet_end": 497.3837403846154, "raw_text": "这种设计还有个更棒的好处", "selected_text": "这种设计还有个更棒的好处", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 233 ], "packet_id": 17, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 492.59575480769234, "end": 495.21575480769235, "text": "既然上半部分只需要用到下半部分生成的KV", "raw_asr_text": "既然上半部分只需要用到下半部分生成的KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 17, "words": [ { "word": "既", "start": 492.5557548076924, "end": 492.7357548076924, "probability": 0.98486328125 }, { "word": "然", "start": 492.7357548076924, "end": 492.83575480769235, "probability": 0.99951171875 }, { "word": "上", "start": 492.83575480769235, "end": 492.9757548076924, "probability": 0.99462890625 }, { "word": "半", "start": 492.9757548076924, "end": 493.1157548076924, "probability": 1.0 }, { "word": "部分", "start": 493.1157548076924, "end": 493.33575480769235, "probability": 0.99951171875 }, { "word": "只", "start": 493.33575480769235, "end": 493.51575480769236, "probability": 0.998046875 }, { "word": "需要", "start": 493.51575480769236, "end": 493.7357548076924, "probability": 0.99658203125 }, { "word": "用", "start": 493.7357548076924, "end": 493.8557548076924, "probability": 0.9990234375 }, { "word": "到", "start": 493.8557548076924, "end": 493.95575480769236, "probability": 0.99365234375 }, { "word": "下", "start": 493.95575480769236, "end": 494.13575480769236, "probability": 0.9990234375 }, { "word": "半", "start": 494.13575480769236, "end": 494.27575480769235, "probability": 1.0 }, { "word": "部分", "start": 494.27575480769235, "end": 494.4957548076924, "probability": 0.99853515625 }, { "word": "生", "start": 494.4957548076924, "end": 494.65575480769235, "probability": 0.99755859375 }, { "word": "成", "start": 494.65575480769235, "end": 494.7957548076924, "probability": 0.9990234375 }, { "word": "的", "start": 494.7957548076924, "end": 494.89575480769236, "probability": 0.99951171875 }, { "word": "K", "start": 494.89575480769236, "end": 495.0557548076924, "probability": 0.98876953125 }, { "word": "V", "start": 495.0557548076924, "end": 495.21575480769235, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.1252976717523566, "compression_ratio": 1.4666666666666666, "no_speech_prob": 2.5849721668746994e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "容易让人联想到它与最初的注意力机制存在关联但它取这个名字仅仅是因为架构包含两种不同类型的解码器下半部分名为自解码器是因为它采用因果资注意力处理序列构建共享的全局KV上半部分名为交叉解码器是因为每个上层都会生成独立的query并对下半部分生成的KV执行交叉注意力计算这种设计还有个更棒的好处既然上半部分只需要用到下半部分生成的KV那么在玉田充阶段严格来说" } ], "id": 234, "review_annotation": "", "row_id": 234, "bounded_row_id": 234, "parent_row_ids": [ 234 ], "parent_row_id": 234, "packet_id": 17, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 468.19575480769237, "packet_end": 497.3837403846154, "raw_text": "既然上半部分只需要用到下半部分生成的KV", "selected_text": "既然上半部分只需要用到下半部分生成的KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 234 ], "packet_id": 17, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 495.21575480769235, "end": 496.57575480769236, "text": "那么在玉田充阶段", "raw_asr_text": "那么在玉田充阶段", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 17, "words": [ { "word": "那么", "start": 495.21575480769235, "end": 495.57575480769236, "probability": 0.990234375 }, { "word": "在", "start": 495.57575480769236, "end": 495.7957548076924, "probability": 0.998046875 }, { "word": "玉", "start": 495.7957548076924, "end": 495.95575480769236, "probability": 0.6806640625 }, { "word": "田", "start": 495.95575480769236, "end": 496.09575480769234, "probability": 0.99365234375 }, { "word": "充", "start": 496.09575480769234, "end": 496.21575480769235, "probability": 0.74169921875 }, { "word": "阶", "start": 496.21575480769235, "end": 496.41575480769234, "probability": 0.97998046875 }, { "word": "段", "start": 496.41575480769234, "end": 496.57575480769236, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1252976717523566, "compression_ratio": 1.4666666666666666, "no_speech_prob": 2.5849721668746994e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "容易让人联想到它与最初的注意力机制存在关联但它取这个名字仅仅是因为架构包含两种不同类型的解码器下半部分名为自解码器是因为它采用因果资注意力处理序列构建共享的全局KV上半部分名为交叉解码器是因为每个上层都会生成独立的query并对下半部分生成的KV执行交叉注意力计算这种设计还有个更棒的好处既然上半部分只需要用到下半部分生成的KV那么在玉田充阶段严格来说" } ], "id": 235, "review_annotation": "", "row_id": 235, "bounded_row_id": 235, "parent_row_ids": [ 235 ], "parent_row_id": 235, "packet_id": 17, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 468.19575480769237, "packet_end": 497.3837403846154, "raw_text": "那么在玉田充阶段", "selected_text": "那么在玉田充阶段", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 235 ], "packet_id": 17, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 496.5837403846154, "end": 497.2837403846154, "text": "严格来说", "raw_asr_text": "严格来说", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "严", "start": 496.56374038461536, "end": 496.8437403846154, "probability": 0.65087890625 }, { "word": "格", "start": 496.8437403846154, "end": 496.9437403846154, "probability": 0.99853515625 }, { "word": "来", "start": 496.9437403846154, "end": 497.0837403846154, "probability": 0.98486328125 }, { "word": "说", "start": 497.0837403846154, "end": 497.2837403846154, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 236, "review_annotation": "", "row_id": 236, "bounded_row_id": 236, "parent_row_ids": [ 236 ], "parent_row_id": 236, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "严格来说", "selected_text": "严格来说", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 236 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 497.2837403846154, "end": 498.8437403846154, "text": "你完全不需要让提示词", "raw_asr_text": "你完全不需要让提示词", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "你", "start": 497.2837403846154, "end": 497.5437403846154, "probability": 0.955078125 }, { "word": "完全", "start": 497.5437403846154, "end": 497.8837403846154, "probability": 0.99658203125 }, { "word": "不", "start": 497.8837403846154, "end": 498.0837403846154, "probability": 0.9951171875 }, { "word": "需要", "start": 498.0837403846154, "end": 498.2837403846154, "probability": 0.998046875 }, { "word": "让", "start": 498.2837403846154, "end": 498.4437403846154, "probability": 0.994140625 }, { "word": "提", "start": 498.4437403846154, "end": 498.5837403846154, "probability": 0.994140625 }, { "word": "示", "start": 498.5837403846154, "end": 498.7237403846154, "probability": 0.99755859375 }, { "word": "词", "start": 498.7237403846154, "end": 498.8437403846154, "probability": 0.9873046875 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 237, "review_annotation": "", "row_id": 237, "bounded_row_id": 237, "parent_row_ids": [ 237 ], "parent_row_id": 237, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "你完全不需要让提示词", "selected_text": "你完全不需要让提示词", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 237 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 498.8437403846154, "end": 500.2837403846154, "text": "经过上半部分的计算", "raw_asr_text": "经过上半部分的计算", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "经", "start": 498.8437403846154, "end": 498.9837403846154, "probability": 0.9931640625 }, { "word": "过", "start": 498.9837403846154, "end": 499.18374038461536, "probability": 0.99951171875 }, { "word": "上", "start": 499.18374038461536, "end": 499.4037403846154, "probability": 0.99755859375 }, { "word": "半", "start": 499.4037403846154, "end": 499.5837403846154, "probability": 0.99951171875 }, { "word": "部分", "start": 499.5837403846154, "end": 499.8437403846154, "probability": 0.998046875 }, { "word": "的", "start": 499.8437403846154, "end": 499.9637403846154, "probability": 0.98681640625 }, { "word": "计", "start": 499.9637403846154, "end": 500.12374038461536, "probability": 0.9990234375 }, { "word": "算", "start": 500.12374038461536, "end": 500.2837403846154, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 238, "review_annotation": "", "row_id": 238, "bounded_row_id": 238, "parent_row_ids": [ 238 ], "parent_row_id": 238, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "经过上半部分的计算", "selected_text": "经过上半部分的计算", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 238 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 500.2837403846154, "end": 501.8437403846154, "text": "你只需要超长提示词", "raw_asr_text": "你只需要超长提示词", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "你", "start": 500.2837403846154, "end": 500.5437403846154, "probability": 0.98974609375 }, { "word": "只", "start": 500.5437403846154, "end": 500.6437403846154, "probability": 0.98095703125 }, { "word": "需要", "start": 500.6437403846154, "end": 500.80374038461537, "probability": 0.59375 }, { "word": "超", "start": 500.80374038461537, "end": 501.0837403846154, "probability": 0.81640625 }, { "word": "长", "start": 501.0837403846154, "end": 501.36374038461537, "probability": 0.8505859375 }, { "word": "提", "start": 501.36374038461537, "end": 501.5437403846154, "probability": 0.9951171875 }, { "word": "示", "start": 501.5437403846154, "end": 501.68374038461536, "probability": 1.0 }, { "word": "词", "start": 501.68374038461536, "end": 501.8437403846154, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 239, "review_annotation": "", "row_id": 239, "bounded_row_id": 239, "parent_row_ids": [ 239 ], "parent_row_id": 239, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "你只需要超长提示词", "selected_text": "你只需要超长提示词", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 239 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 501.8437403846154, "end": 503.3437403846154, "text": "通过自解码器处理", "raw_asr_text": "通过自解码器处理", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "通", "start": 501.8437403846154, "end": 501.9837403846154, "probability": 0.98486328125 }, { "word": "过", "start": 501.9837403846154, "end": 502.1637403846154, "probability": 0.99951171875 }, { "word": "自", "start": 502.1637403846154, "end": 502.42374038461537, "probability": 0.80615234375 }, { "word": "解", "start": 502.42374038461537, "end": 502.62374038461536, "probability": 0.98388671875 }, { "word": "码", "start": 502.62374038461536, "end": 502.80374038461537, "probability": 0.99853515625 }, { "word": "器", "start": 502.80374038461537, "end": 502.9637403846154, "probability": 0.998046875 }, { "word": "处", "start": 502.9637403846154, "end": 503.18374038461536, "probability": 0.998046875 }, { "word": "理", "start": 503.18374038461536, "end": 503.3437403846154, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 240, "review_annotation": "", "row_id": 240, "bounded_row_id": 240, "parent_row_ids": [ 240 ], "parent_row_id": 240, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "通过自解码器处理", "selected_text": "通过自解码器处理", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 240 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 503.3437403846154, "end": 505.2237403846154, "text": "生成共享全局记忆后", "raw_asr_text": "生成共享全局记忆后", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "生", "start": 503.3437403846154, "end": 503.7637403846154, "probability": 0.97802734375 }, { "word": "成", "start": 503.7637403846154, "end": 503.9837403846154, "probability": 0.93896484375 }, { "word": "共", "start": 503.9837403846154, "end": 504.2637403846154, "probability": 0.953125 }, { "word": "享", "start": 504.2637403846154, "end": 504.4637403846154, "probability": 0.99267578125 }, { "word": "全", "start": 504.4637403846154, "end": 504.6637403846154, "probability": 0.9765625 }, { "word": "局", "start": 504.6637403846154, "end": 504.7837403846154, "probability": 0.84912109375 }, { "word": "记", "start": 504.7837403846154, "end": 504.92374038461537, "probability": 0.77783203125 }, { "word": "忆", "start": 504.92374038461537, "end": 505.0037403846154, "probability": 0.9990234375 }, { "word": "后", "start": 505.0037403846154, "end": 505.2237403846154, "probability": 0.99365234375 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 241, "review_annotation": "", "row_id": 241, "bounded_row_id": 241, "parent_row_ids": [ 241 ], "parent_row_id": 241, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "生成共享全局记忆后", "selected_text": "生成共享全局记忆后", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 241 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 505.2237403846154, "end": 506.2637403846154, "text": "就可以停止操作", "raw_asr_text": "就可以停止操作", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "就可以", "start": 505.2237403846154, "end": 505.5037403846154, "probability": 0.978515625 }, { "word": "停", "start": 505.5037403846154, "end": 505.7837403846154, "probability": 0.9990234375 }, { "word": "止", "start": 505.7837403846154, "end": 505.9437403846154, "probability": 0.99951171875 }, { "word": "操", "start": 505.9437403846154, "end": 506.0837403846154, "probability": 0.99951171875 }, { "word": "作", "start": 506.0837403846154, "end": 506.2637403846154, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 242, "review_annotation": "", "row_id": 242, "bounded_row_id": 242, "parent_row_ids": [ 242 ], "parent_row_id": 242, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "就可以停止操作", "selected_text": "就可以停止操作", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 242 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 506.2637403846154, "end": 506.86374038461537, "text": "这意味着", "raw_asr_text": "这意味着", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "这", "start": 506.2637403846154, "end": 506.5037403846154, "probability": 0.98828125 }, { "word": "意", "start": 506.5037403846154, "end": 506.6437403846154, "probability": 0.998046875 }, { "word": "味", "start": 506.6437403846154, "end": 506.7237403846154, "probability": 0.9990234375 }, { "word": "着", "start": 506.7237403846154, "end": 506.86374038461537, "probability": 0.99462890625 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 243, "review_annotation": "", "row_id": 243, "bounded_row_id": 243, "parent_row_ids": [ 243 ], "parent_row_id": 243, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "这意味着", "selected_text": "这意味着", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 243 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 506.86374038461537, "end": 509.1037403846154, "text": "Yoko其实首创了这一思路", "raw_asr_text": "Yoko其实首创了这一思路", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "Y", "start": 506.86374038461537, "end": 507.1037403846154, "probability": 0.77099609375 }, { "word": "oko", "start": 507.1037403846154, "end": 507.2637403846154, "probability": 0.58740234375 }, { "word": "其实", "start": 507.2637403846154, "end": 507.62374038461536, "probability": 0.96142578125 }, { "word": "首", "start": 507.62374038461536, "end": 507.9637403846154, "probability": 0.99560546875 }, { "word": "创", "start": 507.9637403846154, "end": 508.2037403846154, "probability": 0.9990234375 }, { "word": "了", "start": 508.2037403846154, "end": 508.3437403846154, "probability": 0.99609375 }, { "word": "这", "start": 508.3437403846154, "end": 508.4837403846154, "probability": 0.99267578125 }, { "word": "一", "start": 508.4837403846154, "end": 508.7037403846154, "probability": 0.99658203125 }, { "word": "思", "start": 508.7037403846154, "end": 508.9037403846154, "probability": 0.9951171875 }, { "word": "路", "start": 508.9037403846154, "end": 509.1037403846154, "probability": 0.99658203125 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 244, "review_annotation": "", "row_id": 244, "bounded_row_id": 244, "parent_row_ids": [ 244 ], "parent_row_id": 244, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "Yoko其实首创了这一思路", "selected_text": "Yoko其实首创了这一思路", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 244 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 509.1037403846154, "end": 510.8237403846154, "text": "预填充和解码阶段", "raw_asr_text": "预填充和解码阶段", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "预", "start": 509.1037403846154, "end": 509.5437403846154, "probability": 0.94140625 }, { "word": "填", "start": 509.5437403846154, "end": 509.7837403846154, "probability": 1.0 }, { "word": "充", "start": 509.7837403846154, "end": 509.9837403846154, "probability": 0.99951171875 }, { "word": "和", "start": 509.9837403846154, "end": 510.18374038461536, "probability": 0.99560546875 }, { "word": "解", "start": 510.18374038461536, "end": 510.3837403846154, "probability": 0.9990234375 }, { "word": "码", "start": 510.3837403846154, "end": 510.5437403846154, "probability": 1.0 }, { "word": "阶", "start": 510.5437403846154, "end": 510.6637403846154, "probability": 0.99951171875 }, { "word": "段", "start": 510.6637403846154, "end": 510.8237403846154, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 245, "review_annotation": "", "row_id": 245, "bounded_row_id": 245, "parent_row_ids": [ 245 ], "parent_row_id": 245, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "预填充和解码阶段", "selected_text": "预填充和解码阶段", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 245 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 510.8237403846154, "end": 513.0637403846154, "text": "未必要用到同等规模的模型", "raw_asr_text": "未必要用到同等规模的模型", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "未", "start": 510.8237403846154, "end": 511.0837403846154, "probability": 0.99755859375 }, { "word": "必", "start": 511.0837403846154, "end": 511.24374038461536, "probability": 1.0 }, { "word": "要", "start": 511.24374038461536, "end": 511.3837403846154, "probability": 0.99853515625 }, { "word": "用", "start": 511.3837403846154, "end": 511.5037403846154, "probability": 0.99755859375 }, { "word": "到", "start": 511.5037403846154, "end": 511.68374038461536, "probability": 0.9990234375 }, { "word": "同", "start": 511.68374038461536, "end": 512.0237403846154, "probability": 0.99560546875 }, { "word": "等", "start": 512.0237403846154, "end": 512.2437403846154, "probability": 0.9990234375 }, { "word": "规", "start": 512.2437403846154, "end": 512.4437403846154, "probability": 1.0 }, { "word": "模", "start": 512.4437403846154, "end": 512.5637403846154, "probability": 0.99951171875 }, { "word": "的", "start": 512.5637403846154, "end": 512.6637403846154, "probability": 0.9921875 }, { "word": "模", "start": 512.6637403846154, "end": 512.8037403846154, "probability": 0.9990234375 }, { "word": "型", "start": 512.8037403846154, "end": 513.0637403846154, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 246, "review_annotation": "", "row_id": 246, "bounded_row_id": 246, "parent_row_ids": [ 246 ], "parent_row_id": 246, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "未必要用到同等规模的模型", "selected_text": "未必要用到同等规模的模型", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 246 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 513.0637403846154, "end": 514.0837403846153, "text": "当输入规模", "raw_asr_text": "当输入规模", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "当", "start": 513.0637403846154, "end": 513.3637403846154, "probability": 0.9951171875 }, { "word": "输", "start": 513.3637403846154, "end": 513.6237403846154, "probability": 0.99951171875 }, { "word": "入", "start": 513.6237403846154, "end": 513.7637403846154, "probability": 1.0 }, { "word": "规", "start": 513.7637403846154, "end": 513.9237403846154, "probability": 0.9990234375 }, { "word": "模", "start": 513.9237403846154, "end": 514.0837403846153, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 247, "review_annotation": "", "row_id": 247, "bounded_row_id": 247, "parent_row_ids": [ 247 ], "parent_row_id": 247, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "当输入规模", "selected_text": "当输入规模", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 247 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 514.0837403846153, "end": 515.9437403846154, "text": "甚至达到100万tokens时", "raw_asr_text": "甚至达到100万tokens时", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "甚", "start": 514.0837403846153, "end": 514.2637403846154, "probability": 0.99365234375 }, { "word": "至", "start": 514.2637403846154, "end": 514.3837403846154, "probability": 1.0 }, { "word": "达", "start": 514.3837403846154, "end": 514.5437403846154, "probability": 0.99951171875 }, { "word": "到", "start": 514.5437403846154, "end": 514.7237403846154, "probability": 0.99658203125 }, { "word": "100", "start": 514.7237403846154, "end": 515.0637403846154, "probability": 0.94189453125 }, { "word": "万", "start": 515.0637403846154, "end": 515.3237403846153, "probability": 0.98486328125 }, { "word": "t", "start": 515.3237403846153, "end": 515.4837403846154, "probability": 0.61181640625 }, { "word": "ok", "start": 515.4837403846154, "end": 515.5837403846153, "probability": 0.56640625 }, { "word": "ens", "start": 515.5837403846153, "end": 515.7037403846153, "probability": 0.99365234375 }, { "word": "时", "start": 515.7037403846153, "end": 515.9437403846154, "probability": 0.978515625 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 248, "review_annotation": "", "row_id": 248, "bounded_row_id": 248, "parent_row_ids": [ 248 ], "parent_row_id": 248, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "甚至达到100万tokens时", "selected_text": "甚至达到100万tokens时", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 248 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 515.9437403846154, "end": 517.2837403846154, "text": "这个设计思路的优势", "raw_asr_text": "这个设计思路的优势", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "这个", "start": 515.9437403846154, "end": 516.3437403846153, "probability": 0.96044921875 }, { "word": "设", "start": 516.3437403846153, "end": 516.4837403846154, "probability": 0.99951171875 }, { "word": "计", "start": 516.4837403846154, "end": 516.6037403846154, "probability": 1.0 }, { "word": "思", "start": 516.6037403846154, "end": 516.7637403846154, "probability": 0.9970703125 }, { "word": "路", "start": 516.7637403846154, "end": 516.8637403846154, "probability": 0.99365234375 }, { "word": "的", "start": 516.8637403846154, "end": 516.9837403846154, "probability": 0.9951171875 }, { "word": "优", "start": 516.9837403846154, "end": 517.1237403846154, "probability": 0.9970703125 }, { "word": "势", "start": 517.1237403846154, "end": 517.2837403846154, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 249, "review_annotation": "", "row_id": 249, "bounded_row_id": 249, "parent_row_ids": [ 249 ], "parent_row_id": 249, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "这个设计思路的优势", "selected_text": "这个设计思路的优势", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 249 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 517.2837403846154, "end": 518.2237403846154, "text": "就十分突出了", "raw_asr_text": "就十分突出了", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "就", "start": 517.2837403846154, "end": 517.4037403846154, "probability": 0.99365234375 }, { "word": "十", "start": 517.4037403846154, "end": 517.5637403846154, "probability": 0.9931640625 }, { "word": "分", "start": 517.5637403846154, "end": 517.7037403846153, "probability": 0.99951171875 }, { "word": "突", "start": 517.7037403846153, "end": 517.8637403846154, "probability": 0.99658203125 }, { "word": "出", "start": 517.8637403846154, "end": 518.1037403846154, "probability": 0.9990234375 }, { "word": "了", "start": 518.1037403846154, "end": 518.2237403846154, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 250, "review_annotation": "", "row_id": 250, "bounded_row_id": 250, "parent_row_ids": [ 250 ], "parent_row_id": 250, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "就十分突出了", "selected_text": "就十分突出了", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 250 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 518.2237403846154, "end": 520.4837403846154, "text": "因此Deepseek借鉴了这一构想", "raw_asr_text": "因此Deepseek借鉴了这一构想", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "因", "start": 518.2237403846154, "end": 518.5637403846154, "probability": 0.99658203125 }, { "word": "此", "start": 518.5637403846154, "end": 518.7437403846154, "probability": 1.0 }, { "word": "De", "start": 518.7437403846154, "end": 518.9437403846154, "probability": 0.82421875 }, { "word": "ep", "start": 518.9437403846154, "end": 519.0437403846154, "probability": 0.95556640625 }, { "word": "se", "start": 519.0437403846154, "end": 519.1637403846154, "probability": 0.401611328125 }, { "word": "ek", "start": 519.1637403846154, "end": 519.3037403846154, "probability": 0.86767578125 }, { "word": "借", "start": 519.3037403846154, "end": 519.5237403846154, "probability": 0.990234375 }, { "word": "鉴", "start": 519.5237403846154, "end": 519.6437403846154, "probability": 0.99609375 }, { "word": "了", "start": 519.6437403846154, "end": 519.8237403846153, "probability": 0.99658203125 }, { "word": "这", "start": 519.8237403846153, "end": 520.0037403846154, "probability": 0.998046875 }, { "word": "一", "start": 520.0037403846154, "end": 520.1437403846154, "probability": 0.9970703125 }, { "word": "构", "start": 520.1437403846154, "end": 520.2837403846154, "probability": 0.9931640625 }, { "word": "想", "start": 520.2837403846154, "end": 520.4837403846154, "probability": 0.99609375 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 251, "review_annotation": "", "row_id": 251, "bounded_row_id": 251, "parent_row_ids": [ 251 ], "parent_row_id": 251, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "因此Deepseek借鉴了这一构想", "selected_text": "因此Deepseek借鉴了这一构想", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 251 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 520.4837403846154, "end": 521.8637403846154, "text": "并对其进行适配", "raw_asr_text": "并对其进行适配", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "并", "start": 520.4837403846154, "end": 520.8037403846154, "probability": 0.994140625 }, { "word": "对", "start": 520.8037403846154, "end": 520.9637403846153, "probability": 0.9990234375 }, { "word": "其", "start": 520.9637403846153, "end": 521.1637403846154, "probability": 0.99609375 }, { "word": "进", "start": 521.1637403846154, "end": 521.3637403846154, "probability": 0.99853515625 }, { "word": "行", "start": 521.3637403846154, "end": 521.5237403846154, "probability": 0.99951171875 }, { "word": "适", "start": 521.5237403846154, "end": 521.6837403846154, "probability": 0.982421875 }, { "word": "配", "start": 521.6837403846154, "end": 521.8637403846154, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 252, "review_annotation": "", "row_id": 252, "bounded_row_id": 252, "parent_row_ids": [ 252 ], "parent_row_id": 252, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "并对其进行适配", "selected_text": "并对其进行适配", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 252 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 521.8637403846154, "end": 524.0237403846154, "text": "使其能兼容V4的现有架构", "raw_asr_text": "使其能兼容V4的现有架构", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 18, "words": [ { "word": "使", "start": 521.8637403846154, "end": 522.1637403846154, "probability": 0.9931640625 }, { "word": "其", "start": 522.1637403846154, "end": 522.3237403846153, "probability": 0.99951171875 }, { "word": "能", "start": 522.3237403846153, "end": 522.4637403846153, "probability": 0.99853515625 }, { "word": "兼", "start": 522.4637403846153, "end": 522.6637403846154, "probability": 0.998046875 }, { "word": "容", "start": 522.6637403846154, "end": 522.8237403846153, "probability": 0.99951171875 }, { "word": "V", "start": 522.8237403846153, "end": 523.0237403846154, "probability": 0.6767578125 }, { "word": "4", "start": 523.0237403846154, "end": 523.2437403846154, "probability": 0.990234375 }, { "word": "的", "start": 523.2437403846154, "end": 523.3837403846154, "probability": 0.9970703125 }, { "word": "现", "start": 523.3837403846154, "end": 523.5437403846154, "probability": 0.99609375 }, { "word": "有", "start": 523.5437403846154, "end": 523.6637403846154, "probability": 0.99951171875 }, { "word": "架", "start": 523.6637403846154, "end": 523.8037403846154, "probability": 0.998046875 }, { "word": "构", "start": 523.8037403846154, "end": 524.0237403846154, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.09259306165907118, "compression_ratio": 1.2574525745257452, "no_speech_prob": 4.9845030880568686e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在预填充阶段严格来说你完全不需要让提示词经过上半部分的计算你只需要超长提示词通过自解码器处理生成共享全局记忆后就可以停止操作这意味着Yoko其实首创了这一思路预填充和解码阶段未必要用到同等规模的模型当输入规模甚至达到100万tokens时这个设计思路的优势就十分突出了因此Deepseek借鉴了这一构想并对其进行适配使其能兼容V4的现有架构 DeepSeek不" } ], "id": 253, "review_annotation": "", "row_id": 253, "bounded_row_id": 253, "parent_row_ids": [ 253 ], "parent_row_id": 253, "packet_id": 18, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 495.7837403846154, "packet_end": 524.9717259615385, "raw_text": "使其能兼容V4的现有架构", "selected_text": "使其能兼容V4的现有架构", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 253 ], "packet_id": 18, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 524.1717259615385, "end": 527.1517259615385, "text": "Deepseek不再把下半部分称为自解码器", "raw_asr_text": "Deepseek不再把下半部分称为自解码器", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 19, "words": [ { "word": "De", "start": 524.0317259615385, "end": 524.3317259615386, "probability": 0.61572265625 }, { "word": "ep", "start": 524.3317259615386, "end": 524.4317259615385, "probability": 0.9609375 }, { "word": "se", "start": 524.4317259615385, "end": 524.5717259615386, "probability": 0.312255859375 }, { "word": "ek", "start": 524.5717259615386, "end": 524.7117259615386, "probability": 0.9248046875 }, { "word": "不", "start": 524.7117259615386, "end": 524.9717259615386, "probability": 0.96826171875 }, { "word": "再", "start": 524.9717259615386, "end": 525.1517259615385, "probability": 0.99560546875 }, { "word": "把", "start": 525.1517259615385, "end": 525.3117259615386, "probability": 0.994140625 }, { "word": "下", "start": 525.3117259615386, "end": 525.5117259615386, "probability": 0.998046875 }, { "word": "半", "start": 525.5117259615386, "end": 525.6517259615385, "probability": 0.99951171875 }, { "word": "部分", "start": 525.6517259615385, "end": 525.9117259615385, "probability": 0.98681640625 }, { "word": "称", "start": 525.9117259615385, "end": 526.2317259615386, "probability": 0.9951171875 }, { "word": "为", "start": 526.2317259615386, "end": 526.4117259615385, "probability": 0.99658203125 }, { "word": "自", "start": 526.4117259615385, "end": 526.6117259615386, "probability": 0.7998046875 }, { "word": "解", "start": 526.6117259615386, "end": 526.7717259615386, "probability": 0.92578125 }, { "word": "码", "start": 526.7717259615386, "end": 526.9917259615386, "probability": 0.99853515625 }, { "word": "器", "start": 526.9917259615386, "end": 527.1517259615385, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.10092281542326274, "compression_ratio": 1.4190476190476191, "no_speech_prob": 4.394258568130027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "现有架构Deepseek不再把下半部分称为自解码器而是称之为因果编码器它没有像Yoko那样让上半部分共享同一份全局KV而是让每个解码器层从最终编码器表示派生各自的全局KV所以从功能上讲Yoko在KV缓存共享方面要激进的多而Deepseek的因果编码器仍会在模型前半部分之后给出这一个最终表示但它并没有把这个表示变成所有层都必须共享的同一层" } ], "id": 254, "review_annotation": "", "row_id": 254, "bounded_row_id": 254, "parent_row_ids": [ 254 ], "parent_row_id": 254, "packet_id": 19, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 523.3717259615386, "packet_end": 552.5597115384616, "raw_text": "Deepseek不再把下半部分称为自解码器", "selected_text": "Deepseek不再把下半部分称为自解码器", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 254 ], "packet_id": 19, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 527.1517259615385, "end": 529.1917259615386, "text": "而是称之为因果编码器", "raw_asr_text": "而是称之为因果编码器", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 19, "words": [ { "word": "而", "start": 527.1517259615385, "end": 527.4317259615385, "probability": 0.97607421875 }, { "word": "是", "start": 527.4317259615385, "end": 527.5717259615386, "probability": 0.9990234375 }, { "word": "称", "start": 527.5717259615386, "end": 527.7717259615386, "probability": 1.0 }, { "word": "之", "start": 527.7717259615386, "end": 527.9117259615385, "probability": 0.9970703125 }, { "word": "为", "start": 527.9117259615385, "end": 528.2117259615386, "probability": 0.998046875 }, { "word": "因", "start": 528.2117259615386, "end": 528.4317259615385, "probability": 0.98095703125 }, { "word": "果", "start": 528.4317259615385, "end": 528.6717259615385, "probability": 1.0 }, { "word": "编", "start": 528.6717259615385, "end": 528.8517259615386, "probability": 0.99365234375 }, { "word": "码", "start": 528.8517259615386, "end": 529.0317259615385, "probability": 1.0 }, { "word": "器", "start": 529.0317259615385, "end": 529.1917259615386, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10092281542326274, "compression_ratio": 1.4190476190476191, "no_speech_prob": 4.394258568130027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "现有架构Deepseek不再把下半部分称为自解码器而是称之为因果编码器它没有像Yoko那样让上半部分共享同一份全局KV而是让每个解码器层从最终编码器表示派生各自的全局KV所以从功能上讲Yoko在KV缓存共享方面要激进的多而Deepseek的因果编码器仍会在模型前半部分之后给出这一个最终表示但它并没有把这个表示变成所有层都必须共享的同一层" } ], "id": 255, "review_annotation": "", "row_id": 255, "bounded_row_id": 255, "parent_row_ids": [ 255 ], "parent_row_id": 255, "packet_id": 19, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 523.3717259615386, "packet_end": 552.5597115384616, "raw_text": "而是称之为因果编码器", "selected_text": "而是称之为因果编码器", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 255 ], "packet_id": 19, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 529.1917259615386, "end": 533.3317259615386, "text": "它没有像Yoko那样让上半部分共享同一份全局KV", "raw_asr_text": "它没有像Yoko那样让上半部分共享同一份全局KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 19, "words": [ { "word": "它", "start": 529.1917259615386, "end": 529.3717259615386, "probability": 0.76611328125 }, { "word": "没有", "start": 529.3717259615386, "end": 529.6117259615386, "probability": 0.99853515625 }, { "word": "像", "start": 529.6117259615386, "end": 529.8517259615386, "probability": 0.99853515625 }, { "word": "Y", "start": 529.8517259615386, "end": 530.0317259615385, "probability": 0.76953125 }, { "word": "oko", "start": 530.0317259615385, "end": 530.1717259615385, "probability": 0.84423828125 }, { "word": "那", "start": 530.1717259615385, "end": 530.3317259615386, "probability": 0.99609375 }, { "word": "样", "start": 530.3317259615386, "end": 530.4717259615386, "probability": 0.99951171875 }, { "word": "让", "start": 530.4717259615386, "end": 530.6717259615385, "probability": 0.99658203125 }, { "word": "上", "start": 530.6717259615385, "end": 530.9717259615386, "probability": 0.998046875 }, { "word": "半", "start": 530.9717259615386, "end": 531.1717259615385, "probability": 1.0 }, { "word": "部分", "start": 531.1717259615385, "end": 531.4517259615386, "probability": 0.9990234375 }, { "word": "共", "start": 531.4517259615386, "end": 531.7117259615386, "probability": 0.99755859375 }, { "word": "享", "start": 531.7117259615386, "end": 532.0517259615385, "probability": 0.9990234375 }, { "word": "同", "start": 532.0517259615385, "end": 532.3517259615386, "probability": 0.99609375 }, { "word": "一", "start": 532.3517259615386, "end": 532.5717259615386, "probability": 0.99951171875 }, { "word": "份", "start": 532.5717259615386, "end": 532.7117259615386, "probability": 0.9970703125 }, { "word": "全", "start": 532.7117259615386, "end": 532.9317259615385, "probability": 0.9970703125 }, { "word": "局", "start": 532.9317259615385, "end": 533.0317259615385, "probability": 0.61865234375 }, { "word": "K", "start": 533.0317259615385, "end": 533.1717259615385, "probability": 0.783203125 }, { "word": "V", "start": 533.1717259615385, "end": 533.3317259615386, "probability": 0.994140625 } ], "asr_metrics": { "avg_logprob": -0.10092281542326274, "compression_ratio": 1.4190476190476191, "no_speech_prob": 4.394258568130027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "现有架构Deepseek不再把下半部分称为自解码器而是称之为因果编码器它没有像Yoko那样让上半部分共享同一份全局KV而是让每个解码器层从最终编码器表示派生各自的全局KV所以从功能上讲Yoko在KV缓存共享方面要激进的多而Deepseek的因果编码器仍会在模型前半部分之后给出这一个最终表示但它并没有把这个表示变成所有层都必须共享的同一层" } ], "id": 256, "review_annotation": "", "row_id": 256, "bounded_row_id": 256, "parent_row_ids": [ 256 ], "parent_row_id": 256, "packet_id": 19, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 523.3717259615386, "packet_end": 552.5597115384616, "raw_text": "它没有像Yoko那样让上半部分共享同一份全局KV", "selected_text": "它没有像Yoko那样让上半部分共享同一份全局KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 256 ], "packet_id": 19, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 533.3317259615386, "end": 539.0517259615385, "text": "而是让每个解码器层从最终编码器表示派生各自的全局KV", "raw_asr_text": "而是让每个解码器层从最终编码器表示派生各自的全局KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 19, "words": [ { "word": "而", "start": 533.3317259615386, "end": 533.6517259615385, "probability": 0.99462890625 }, { "word": "是", "start": 533.6517259615385, "end": 533.8317259615386, "probability": 1.0 }, { "word": "让", "start": 533.8317259615386, "end": 534.0317259615385, "probability": 0.99853515625 }, { "word": "每", "start": 534.0317259615385, "end": 534.3517259615386, "probability": 0.998046875 }, { "word": "个", "start": 534.3517259615386, "end": 534.5917259615386, "probability": 0.99951171875 }, { "word": "解", "start": 534.5917259615386, "end": 534.7917259615385, "probability": 0.998046875 }, { "word": "码", "start": 534.7917259615385, "end": 534.9717259615386, "probability": 1.0 }, { "word": "器", "start": 534.9717259615386, "end": 535.1517259615385, "probability": 0.99951171875 }, { "word": "层", "start": 535.1517259615385, "end": 535.3317259615386, "probability": 0.99609375 }, { "word": "从", "start": 535.3317259615386, "end": 535.5517259615385, "probability": 0.998046875 }, { "word": "最", "start": 535.5517259615385, "end": 535.8317259615386, "probability": 1.0 }, { "word": "终", "start": 535.8317259615386, "end": 536.0517259615385, "probability": 0.994140625 }, { "word": "编", "start": 536.0517259615385, "end": 536.2917259615385, "probability": 0.99951171875 }, { "word": "码", "start": 536.2917259615385, "end": 536.4917259615386, "probability": 1.0 }, { "word": "器", "start": 536.4917259615386, "end": 536.6517259615385, "probability": 1.0 }, { "word": "表示", "start": 536.6517259615385, "end": 537.0117259615386, "probability": 0.982421875 }, { "word": "派", "start": 537.0117259615386, "end": 537.4717259615386, "probability": 0.99267578125 }, { "word": "生", "start": 537.4717259615386, "end": 537.6917259615386, "probability": 0.689453125 }, { "word": "各", "start": 537.6917259615386, "end": 537.9517259615386, "probability": 0.998046875 }, { "word": "自", "start": 537.9517259615386, "end": 538.1317259615386, "probability": 0.99951171875 }, { "word": "的", "start": 538.1317259615386, "end": 538.2517259615386, "probability": 0.99560546875 }, { "word": "全", "start": 538.2517259615386, "end": 538.5317259615385, "probability": 0.99951171875 }, { "word": "局", "start": 538.5317259615385, "end": 538.7117259615386, "probability": 0.998046875 }, { "word": "K", "start": 538.7117259615386, "end": 538.8517259615386, "probability": 0.99609375 }, { "word": "V", "start": 538.8517259615386, "end": 539.0517259615385, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10092281542326274, "compression_ratio": 1.4190476190476191, "no_speech_prob": 4.394258568130027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "现有架构Deepseek不再把下半部分称为自解码器而是称之为因果编码器它没有像Yoko那样让上半部分共享同一份全局KV而是让每个解码器层从最终编码器表示派生各自的全局KV所以从功能上讲Yoko在KV缓存共享方面要激进的多而Deepseek的因果编码器仍会在模型前半部分之后给出这一个最终表示但它并没有把这个表示变成所有层都必须共享的同一层" } ], "id": 257, "review_annotation": "", "row_id": 257, "bounded_row_id": 257, "parent_row_ids": [ 257 ], "parent_row_id": 257, "packet_id": 19, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 523.3717259615386, "packet_end": 552.5597115384616, "raw_text": "而是让每个解码器层从最终编码器表示派生各自的全局KV", "selected_text": "而是让每个解码器层从最终编码器表示派生各自的全局KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 257 ], "packet_id": 19, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 539.0517259615385, "end": 540.3517259615386, "text": "所以从功能上讲", "raw_asr_text": "所以从功能上讲", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 19, "words": [ { "word": "所以", "start": 539.0517259615385, "end": 539.4517259615386, "probability": 0.9970703125 }, { "word": "从", "start": 539.4517259615386, "end": 539.6917259615386, "probability": 0.99951171875 }, { "word": "功", "start": 539.6917259615386, "end": 539.8717259615386, "probability": 0.9970703125 }, { "word": "能", "start": 539.8717259615386, "end": 540.0117259615386, "probability": 0.99951171875 }, { "word": "上", "start": 540.0117259615386, "end": 540.1717259615385, "probability": 1.0 }, { "word": "讲", "start": 540.1717259615385, "end": 540.3517259615386, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.10092281542326274, "compression_ratio": 1.4190476190476191, "no_speech_prob": 4.394258568130027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "现有架构Deepseek不再把下半部分称为自解码器而是称之为因果编码器它没有像Yoko那样让上半部分共享同一份全局KV而是让每个解码器层从最终编码器表示派生各自的全局KV所以从功能上讲Yoko在KV缓存共享方面要激进的多而Deepseek的因果编码器仍会在模型前半部分之后给出这一个最终表示但它并没有把这个表示变成所有层都必须共享的同一层" } ], "id": 258, "review_annotation": "", "row_id": 258, "bounded_row_id": 258, "parent_row_ids": [ 258 ], "parent_row_id": 258, "packet_id": 19, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 523.3717259615386, "packet_end": 552.5597115384616, "raw_text": "所以从功能上讲", "selected_text": "所以从功能上讲", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 258 ], "packet_id": 19, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 540.3517259615386, "end": 543.2917259615385, "text": "Yoko在KV缓存共享方面要激进的多", "raw_asr_text": "Yoko在KV缓存共享方面要激进的多", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 19, "words": [ { "word": "Y", "start": 540.3517259615386, "end": 540.5917259615386, "probability": 0.95947265625 }, { "word": "oko", "start": 540.5917259615386, "end": 540.7717259615386, "probability": 0.9990234375 }, { "word": "在", "start": 540.7717259615386, "end": 541.0117259615386, "probability": 0.99658203125 }, { "word": "K", "start": 541.0117259615386, "end": 541.2317259615386, "probability": 0.9892578125 }, { "word": "V", "start": 541.2317259615386, "end": 541.3917259615386, "probability": 0.99951171875 }, { "word": "缓", "start": 541.3917259615386, "end": 541.5517259615385, "probability": 0.9990234375 }, { "word": "存", "start": 541.5517259615385, "end": 541.7517259615386, "probability": 0.9990234375 }, { "word": "共", "start": 541.7517259615386, "end": 541.9117259615385, "probability": 0.99560546875 }, { "word": "享", "start": 541.9117259615385, "end": 542.0917259615386, "probability": 0.99951171875 }, { "word": "方面", "start": 542.0917259615386, "end": 542.4117259615385, "probability": 0.9970703125 }, { "word": "要", "start": 542.4117259615385, "end": 542.6917259615386, "probability": 0.99658203125 }, { "word": "激", "start": 542.6917259615386, "end": 542.8717259615386, "probability": 0.765625 }, { "word": "进", "start": 542.8717259615386, "end": 543.0117259615386, "probability": 0.986328125 }, { "word": "的", "start": 543.0117259615386, "end": 543.1117259615386, "probability": 0.85107421875 }, { "word": "多", "start": 543.1117259615386, "end": 543.2917259615385, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.10092281542326274, "compression_ratio": 1.4190476190476191, "no_speech_prob": 4.394258568130027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "现有架构Deepseek不再把下半部分称为自解码器而是称之为因果编码器它没有像Yoko那样让上半部分共享同一份全局KV而是让每个解码器层从最终编码器表示派生各自的全局KV所以从功能上讲Yoko在KV缓存共享方面要激进的多而Deepseek的因果编码器仍会在模型前半部分之后给出这一个最终表示但它并没有把这个表示变成所有层都必须共享的同一层" } ], "id": 259, "review_annotation": "", "row_id": 259, "bounded_row_id": 259, "parent_row_ids": [ 259 ], "parent_row_id": 259, "packet_id": 19, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 523.3717259615386, "packet_end": 552.5597115384616, "raw_text": "Yoko在KV缓存共享方面要激进的多", "selected_text": "Yoko在KV缓存共享方面要激进的多", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 259 ], "packet_id": 19, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 543.2917259615385, "end": 545.3117259615386, "text": "而Deepseek的因果编码器", "raw_asr_text": "而Deepseek的因果编码器", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 19, "words": [ { "word": "而", "start": 543.2917259615385, "end": 543.7117259615386, "probability": 0.99853515625 }, { "word": "De", "start": 543.7117259615386, "end": 543.9317259615385, "probability": 0.9951171875 }, { "word": "ep", "start": 543.9317259615385, "end": 544.0517259615385, "probability": 0.9990234375 }, { "word": "se", "start": 544.0517259615385, "end": 544.1917259615386, "probability": 0.96142578125 }, { "word": "ek", "start": 544.1917259615386, "end": 544.3117259615386, "probability": 1.0 }, { "word": "的", "start": 544.3117259615386, "end": 544.5117259615386, "probability": 0.99267578125 }, { "word": "因", "start": 544.5117259615386, "end": 544.7117259615386, "probability": 0.9990234375 }, { "word": "果", "start": 544.7117259615386, "end": 544.8517259615386, "probability": 1.0 }, { "word": "编", "start": 544.8517259615386, "end": 544.9717259615386, "probability": 0.9990234375 }, { "word": "码", "start": 544.9717259615386, "end": 545.1117259615386, "probability": 1.0 }, { "word": "器", "start": 545.1117259615386, "end": 545.3117259615386, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10092281542326274, "compression_ratio": 1.4190476190476191, "no_speech_prob": 4.394258568130027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "现有架构Deepseek不再把下半部分称为自解码器而是称之为因果编码器它没有像Yoko那样让上半部分共享同一份全局KV而是让每个解码器层从最终编码器表示派生各自的全局KV所以从功能上讲Yoko在KV缓存共享方面要激进的多而Deepseek的因果编码器仍会在模型前半部分之后给出这一个最终表示但它并没有把这个表示变成所有层都必须共享的同一层" } ], "id": 260, "review_annotation": "", "row_id": 260, "bounded_row_id": 260, "parent_row_ids": [ 260 ], "parent_row_id": 260, "packet_id": 19, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 523.3717259615386, "packet_end": 552.5597115384616, "raw_text": "而Deepseek的因果编码器", "selected_text": "而Deepseek的因果编码器", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 260 ], "packet_id": 19, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 545.3117259615386, "end": 548.4317259615385, "text": "仍会在模型前半部分之后给出这一个最终表示", "raw_asr_text": "仍会在模型前半部分之后给出这一个最终表示", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 19, "words": [ { "word": "仍", "start": 545.3117259615386, "end": 545.5717259615386, "probability": 0.966796875 }, { "word": "会", "start": 545.5717259615386, "end": 545.7117259615386, "probability": 0.955078125 }, { "word": "在", "start": 545.7117259615386, "end": 545.8717259615386, "probability": 0.99951171875 }, { "word": "模", "start": 545.8717259615386, "end": 546.0117259615386, "probability": 0.998046875 }, { "word": "型", "start": 546.0117259615386, "end": 546.1917259615386, "probability": 0.99951171875 }, { "word": "前", "start": 546.1917259615386, "end": 546.3917259615386, "probability": 0.99658203125 }, { "word": "半", "start": 546.3917259615386, "end": 546.5317259615385, "probability": 0.99951171875 }, { "word": "部分", "start": 546.5317259615385, "end": 546.7517259615386, "probability": 0.9990234375 }, { "word": "之", "start": 546.7517259615386, "end": 546.9317259615385, "probability": 0.99609375 }, { "word": "后", "start": 546.9317259615385, "end": 547.0517259615385, "probability": 1.0 }, { "word": "给", "start": 547.0517259615385, "end": 547.2117259615386, "probability": 0.99755859375 }, { "word": "出", "start": 547.2117259615386, "end": 547.3717259615386, "probability": 1.0 }, { "word": "这", "start": 547.3717259615386, "end": 547.5317259615385, "probability": 0.95361328125 }, { "word": "一个", "start": 547.5317259615385, "end": 547.7717259615386, "probability": 0.998046875 }, { "word": "最", "start": 547.7717259615386, "end": 547.9917259615386, "probability": 0.9990234375 }, { "word": "终", "start": 547.9917259615386, "end": 548.1317259615386, "probability": 1.0 }, { "word": "表示", "start": 548.1317259615386, "end": 548.4317259615385, "probability": 0.97412109375 } ], "asr_metrics": { "avg_logprob": -0.10092281542326274, "compression_ratio": 1.4190476190476191, "no_speech_prob": 4.394258568130027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "现有架构Deepseek不再把下半部分称为自解码器而是称之为因果编码器它没有像Yoko那样让上半部分共享同一份全局KV而是让每个解码器层从最终编码器表示派生各自的全局KV所以从功能上讲Yoko在KV缓存共享方面要激进的多而Deepseek的因果编码器仍会在模型前半部分之后给出这一个最终表示但它并没有把这个表示变成所有层都必须共享的同一层" } ], "id": 261, "review_annotation": "", "row_id": 261, "bounded_row_id": 261, "parent_row_ids": [ 261 ], "parent_row_id": 261, "packet_id": 19, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 523.3717259615386, "packet_end": 552.5597115384616, "raw_text": "仍会在模型前半部分之后给出这一个最终表示", "selected_text": "仍会在模型前半部分之后给出这一个最终表示", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 261 ], "packet_id": 19, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 548.4317259615385, "end": 551.7597115384616, "text": "但它并没有把这个表示变成所有层都必须共享的同一层", "raw_asr_text": "但它并没有把这个表示变成所有层都必须共享的同一层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 19, "words": [ { "word": "但", "start": 548.4317259615385, "end": 548.8317259615386, "probability": 0.99609375 }, { "word": "它", "start": 548.8317259615386, "end": 549.0117259615386, "probability": 0.876953125 }, { "word": "并", "start": 549.0117259615386, "end": 549.2517259615386, "probability": 1.0 }, { "word": "没有", "start": 549.2517259615386, "end": 549.4717259615386, "probability": 0.99951171875 }, { "word": "把", "start": 549.4717259615386, "end": 549.6517259615385, "probability": 0.99951171875 }, { "word": "这个", "start": 549.6517259615385, "end": 549.8717259615386, "probability": 0.99658203125 }, { "word": "表示", "start": 549.8717259615386, "end": 550.2117259615386, "probability": 0.994140625 }, { "word": "变", "start": 550.2117259615386, "end": 550.5117259615386, "probability": 0.99951171875 }, { "word": "成", "start": 550.5117259615386, "end": 550.6517259615385, "probability": 0.99951171875 }, { "word": "所有", "start": 550.6517259615385, "end": 551.0117259615386, "probability": 0.99609375 }, { "word": "层", "start": 551.0117259615386, "end": 551.2517259615386, "probability": 1.0 }, { "word": "都", "start": 551.2517259615386, "end": 551.4117259615385, "probability": 0.99853515625 }, { "word": "必", "start": 551.4117259615385, "end": 551.6517259615385, "probability": 0.99951171875 }, { "word": "须", "start": 551.6517259615385, "end": 551.8917259615386, "probability": 0.9990234375 }, { "word": "共", "start": 551.8917259615386, "end": 552.0717259615386, "probability": 0.99951171875 }, { "word": "享", "start": 552.0717259615386, "end": 552.2917259615385, "probability": 1.0 }, { "word": "的", "start": 552.2917259615385, "end": 552.4117259615385, "probability": 0.921875 }, { "word": "同", "start": 552.4117259615385, "end": 552.5317259615385, "probability": 0.9326171875 }, { "word": "一", "start": 552.5317259615385, "end": 552.5317259615385, "probability": 0.36865234375 }, { "word": "层", "start": 552.5317259615385, "end": 552.5317259615385, "probability": 0.7060546875 } ], "asr_metrics": { "avg_logprob": -0.10092281542326274, "compression_ratio": 1.4190476190476191, "no_speech_prob": 4.394258568130027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "现有架构Deepseek不再把下半部分称为自解码器而是称之为因果编码器它没有像Yoko那样让上半部分共享同一份全局KV而是让每个解码器层从最终编码器表示派生各自的全局KV所以从功能上讲Yoko在KV缓存共享方面要激进的多而Deepseek的因果编码器仍会在模型前半部分之后给出这一个最终表示但它并没有把这个表示变成所有层都必须共享的同一层" } ], "id": 262, "review_annotation": "", "row_id": 262, "bounded_row_id": 262, "parent_row_ids": [ 262 ], "parent_row_id": 262, "packet_id": 19, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 523.3717259615386, "packet_end": 552.5597115384616, "raw_text": "但它并没有把这个表示变成所有层都必须共享的同一层", "selected_text": "但它并没有把这个表示变成所有层都必须共享的同一层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 262 ], "packet_id": 19, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 551.7597115384616, "end": 553.5797115384617, "text": "所有层都必须共享的同一份KV缓存", "raw_asr_text": "所有层都必须共享的同一份KV缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 20, "words": [ { "word": "所有", "start": 550.9597115384616, "end": 550.9597115384616, "probability": 0.0965576171875 }, { "word": "层", "start": 550.9597115384616, "end": 551.2597115384616, "probability": 0.9970703125 }, { "word": "都", "start": 551.2597115384616, "end": 551.3997115384617, "probability": 0.98681640625 }, { "word": "必", "start": 551.3997115384617, "end": 551.6597115384617, "probability": 0.998046875 }, { "word": "须", "start": 551.6597115384617, "end": 551.8997115384617, "probability": 0.994140625 }, { "word": "共", "start": 551.8997115384617, "end": 552.0597115384617, "probability": 0.99560546875 }, { "word": "享", "start": 552.0597115384617, "end": 552.2797115384617, "probability": 0.9990234375 }, { "word": "的", "start": 552.2797115384617, "end": 552.4197115384617, "probability": 0.96875 }, { "word": "同", "start": 552.4197115384617, "end": 552.6197115384616, "probability": 0.994140625 }, { "word": "一", "start": 552.6197115384616, "end": 552.8197115384617, "probability": 0.9990234375 }, { "word": "份", "start": 552.8197115384617, "end": 552.9197115384617, "probability": 0.9990234375 }, { "word": "K", "start": 552.9197115384617, "end": 553.1197115384616, "probability": 0.68505859375 }, { "word": "V", "start": 553.1197115384616, "end": 553.2197115384616, "probability": 0.62939453125 }, { "word": "缓", "start": 553.2197115384616, "end": 553.3997115384617, "probability": 0.9970703125 }, { "word": "存", "start": 553.3997115384617, "end": 553.5797115384617, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08723285655283557, "compression_ratio": 1.3487031700288183, "no_speech_prob": 8.208766999473482e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有层都必须共享的同一份KV缓存而是让每个解码器层都对该最终表示进行各自学习到的投影第20层之后的每一层都可以从该最终表示进行投影并通过各自层专属的权重得到不同结果但这种做法为什么可行呢他们的理由是这是模型当前对该token已知信息的一个庞大的学习性表示因为在经过20个Transformer层之后每个token表示都已经经历了相当深度的上下文处理一起吧" } ], "id": 263, "review_annotation": "", "row_id": 263, "bounded_row_id": 263, "parent_row_ids": [ 263 ], "parent_row_id": 263, "packet_id": 20, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 550.9597115384616, "packet_end": 580.1476971153846, "raw_text": "所有层都必须共享的同一份KV缓存", "selected_text": "所有层都必须共享的同一份KV缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 263 ], "packet_id": 20, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 553.5797115384617, "end": 555.7597115384616, "text": "而是让每个解码器层", "raw_asr_text": "而是让每个解码器层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 20, "words": [ { "word": "而", "start": 553.5797115384617, "end": 553.9797115384616, "probability": 0.9716796875 }, { "word": "是", "start": 553.9797115384616, "end": 554.1997115384617, "probability": 0.9990234375 }, { "word": "让", "start": 554.1997115384617, "end": 554.4397115384617, "probability": 0.974609375 }, { "word": "每", "start": 554.4397115384617, "end": 554.7597115384616, "probability": 0.99658203125 }, { "word": "个", "start": 554.7597115384616, "end": 555.0197115384616, "probability": 0.9970703125 }, { "word": "解", "start": 555.0197115384616, "end": 555.1797115384617, "probability": 0.99267578125 }, { "word": "码", "start": 555.1797115384617, "end": 555.3597115384616, "probability": 0.99755859375 }, { "word": "器", "start": 555.3597115384616, "end": 555.4997115384616, "probability": 0.95361328125 }, { "word": "层", "start": 555.4997115384616, "end": 555.7597115384616, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.08723285655283557, "compression_ratio": 1.3487031700288183, "no_speech_prob": 8.208766999473482e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有层都必须共享的同一份KV缓存而是让每个解码器层都对该最终表示进行各自学习到的投影第20层之后的每一层都可以从该最终表示进行投影并通过各自层专属的权重得到不同结果但这种做法为什么可行呢他们的理由是这是模型当前对该token已知信息的一个庞大的学习性表示因为在经过20个Transformer层之后每个token表示都已经经历了相当深度的上下文处理一起吧" } ], "id": 264, "review_annotation": "", "row_id": 264, "bounded_row_id": 264, "parent_row_ids": [ 264 ], "parent_row_id": 264, "packet_id": 20, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 550.9597115384616, "packet_end": 580.1476971153846, "raw_text": "而是让每个解码器层", "selected_text": "而是让每个解码器层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 264 ], "packet_id": 20, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 555.7597115384616, "end": 559.1197115384616, "text": "都对该最终表示进行各自学习到的投影", "raw_asr_text": "都对该最终表示进行各自学习到的投影", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 20, "words": [ { "word": "都", "start": 555.7597115384616, "end": 556.0197115384616, "probability": 0.9912109375 }, { "word": "对", "start": 556.0197115384616, "end": 556.2597115384616, "probability": 0.9990234375 }, { "word": "该", "start": 556.2597115384616, "end": 556.5597115384617, "probability": 0.9990234375 }, { "word": "最", "start": 556.5597115384617, "end": 556.7797115384617, "probability": 0.99853515625 }, { "word": "终", "start": 556.7797115384617, "end": 556.9797115384616, "probability": 0.9990234375 }, { "word": "表示", "start": 556.9797115384616, "end": 557.2997115384617, "probability": 0.98388671875 }, { "word": "进", "start": 557.2997115384617, "end": 557.5397115384617, "probability": 0.998046875 }, { "word": "行", "start": 557.5397115384617, "end": 557.7797115384617, "probability": 0.99951171875 }, { "word": "各", "start": 557.7797115384617, "end": 557.9797115384616, "probability": 0.998046875 }, { "word": "自", "start": 557.9797115384616, "end": 558.1597115384617, "probability": 0.99951171875 }, { "word": "学", "start": 558.1597115384617, "end": 558.3597115384616, "probability": 0.9990234375 }, { "word": "习", "start": 558.3597115384616, "end": 558.5197115384616, "probability": 0.99951171875 }, { "word": "到", "start": 558.5197115384616, "end": 558.6197115384616, "probability": 0.95361328125 }, { "word": "的", "start": 558.6197115384616, "end": 558.7397115384616, "probability": 0.99755859375 }, { "word": "投", "start": 558.7397115384616, "end": 558.8997115384617, "probability": 0.99951171875 }, { "word": "影", "start": 558.8997115384617, "end": 559.1197115384616, "probability": 0.9912109375 } ], "asr_metrics": { "avg_logprob": -0.08723285655283557, "compression_ratio": 1.3487031700288183, "no_speech_prob": 8.208766999473482e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有层都必须共享的同一份KV缓存而是让每个解码器层都对该最终表示进行各自学习到的投影第20层之后的每一层都可以从该最终表示进行投影并通过各自层专属的权重得到不同结果但这种做法为什么可行呢他们的理由是这是模型当前对该token已知信息的一个庞大的学习性表示因为在经过20个Transformer层之后每个token表示都已经经历了相当深度的上下文处理一起吧" } ], "id": 265, "review_annotation": "", "row_id": 265, "bounded_row_id": 265, "parent_row_ids": [ 265 ], "parent_row_id": 265, "packet_id": 20, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 550.9597115384616, "packet_end": 580.1476971153846, "raw_text": "都对该最终表示进行各自学习到的投影", "selected_text": "都对该最终表示进行各自学习到的投影", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 265 ], "packet_id": 20, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 559.1197115384616, "end": 560.5997115384616, "text": "第20层之后的每一层", "raw_asr_text": "第20层之后的每一层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 20, "words": [ { "word": "第", "start": 559.1197115384616, "end": 559.3797115384616, "probability": 0.8955078125 }, { "word": "20", "start": 559.3797115384616, "end": 559.5797115384617, "probability": 0.9951171875 }, { "word": "层", "start": 559.5797115384617, "end": 559.7397115384616, "probability": 1.0 }, { "word": "之", "start": 559.7397115384616, "end": 559.9197115384617, "probability": 0.998046875 }, { "word": "后", "start": 559.9197115384617, "end": 560.0397115384617, "probability": 1.0 }, { "word": "的", "start": 560.0397115384617, "end": 560.1597115384617, "probability": 0.98779296875 }, { "word": "每", "start": 560.1597115384617, "end": 560.2997115384617, "probability": 0.99951171875 }, { "word": "一", "start": 560.2997115384617, "end": 560.4197115384617, "probability": 0.990234375 }, { "word": "层", "start": 560.4197115384617, "end": 560.5997115384616, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08723285655283557, "compression_ratio": 1.3487031700288183, "no_speech_prob": 8.208766999473482e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有层都必须共享的同一份KV缓存而是让每个解码器层都对该最终表示进行各自学习到的投影第20层之后的每一层都可以从该最终表示进行投影并通过各自层专属的权重得到不同结果但这种做法为什么可行呢他们的理由是这是模型当前对该token已知信息的一个庞大的学习性表示因为在经过20个Transformer层之后每个token表示都已经经历了相当深度的上下文处理一起吧" } ], "id": 266, "review_annotation": "", "row_id": 266, "bounded_row_id": 266, "parent_row_ids": [ 266 ], "parent_row_id": 266, "packet_id": 20, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 550.9597115384616, "packet_end": 580.1476971153846, "raw_text": "第20层之后的每一层", "selected_text": "第20层之后的每一层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 266 ], "packet_id": 20, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 560.5997115384616, "end": 562.6397115384616, "text": "都可以从该最终表示进行投影", "raw_asr_text": "都可以从该最终表示进行投影", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 20, "words": [ { "word": "都", "start": 560.5997115384616, "end": 560.7997115384617, "probability": 0.994140625 }, { "word": "可以", "start": 560.7997115384617, "end": 560.9997115384616, "probability": 0.9990234375 }, { "word": "从", "start": 560.9997115384616, "end": 561.1597115384617, "probability": 0.9990234375 }, { "word": "该", "start": 561.1597115384617, "end": 561.3597115384616, "probability": 0.998046875 }, { "word": "最", "start": 561.3597115384616, "end": 561.5397115384617, "probability": 0.99951171875 }, { "word": "终", "start": 561.5397115384617, "end": 561.6397115384616, "probability": 1.0 }, { "word": "表示", "start": 561.6397115384616, "end": 561.9797115384616, "probability": 0.9990234375 }, { "word": "进", "start": 561.9797115384616, "end": 562.1997115384617, "probability": 0.998046875 }, { "word": "行", "start": 562.1997115384617, "end": 562.3197115384617, "probability": 0.99951171875 }, { "word": "投", "start": 562.3197115384617, "end": 562.4397115384617, "probability": 1.0 }, { "word": "影", "start": 562.4397115384617, "end": 562.6397115384616, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08723285655283557, "compression_ratio": 1.3487031700288183, "no_speech_prob": 8.208766999473482e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有层都必须共享的同一份KV缓存而是让每个解码器层都对该最终表示进行各自学习到的投影第20层之后的每一层都可以从该最终表示进行投影并通过各自层专属的权重得到不同结果但这种做法为什么可行呢他们的理由是这是模型当前对该token已知信息的一个庞大的学习性表示因为在经过20个Transformer层之后每个token表示都已经经历了相当深度的上下文处理一起吧" } ], "id": 267, "review_annotation": "", "row_id": 267, "bounded_row_id": 267, "parent_row_ids": [ 267 ], "parent_row_id": 267, "packet_id": 20, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 550.9597115384616, "packet_end": 580.1476971153846, "raw_text": "都可以从该最终表示进行投影", "selected_text": "都可以从该最终表示进行投影", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 267 ], "packet_id": 20, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 562.6397115384616, "end": 564.4997115384616, "text": "并通过各自层专属的权重", "raw_asr_text": "并通过各自层专属的权重", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 20, "words": [ { "word": "并", "start": 562.6397115384616, "end": 562.8997115384617, "probability": 0.994140625 }, { "word": "通", "start": 562.8997115384617, "end": 562.9997115384616, "probability": 0.99658203125 }, { "word": "过", "start": 562.9997115384616, "end": 563.1597115384617, "probability": 0.99951171875 }, { "word": "各", "start": 563.1597115384617, "end": 563.2997115384617, "probability": 0.9990234375 }, { "word": "自", "start": 563.2997115384617, "end": 563.4797115384616, "probability": 0.986328125 }, { "word": "层", "start": 563.4797115384616, "end": 563.6797115384617, "probability": 0.998046875 }, { "word": "专", "start": 563.6797115384617, "end": 563.8997115384617, "probability": 0.998046875 }, { "word": "属", "start": 563.8997115384617, "end": 564.0397115384617, "probability": 0.998046875 }, { "word": "的", "start": 564.0397115384617, "end": 564.1397115384616, "probability": 0.9990234375 }, { "word": "权", "start": 564.1397115384616, "end": 564.3397115384616, "probability": 0.99951171875 }, { "word": "重", "start": 564.3397115384616, "end": 564.4997115384616, "probability": 0.9658203125 } ], "asr_metrics": { "avg_logprob": -0.08723285655283557, "compression_ratio": 1.3487031700288183, "no_speech_prob": 8.208766999473482e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有层都必须共享的同一份KV缓存而是让每个解码器层都对该最终表示进行各自学习到的投影第20层之后的每一层都可以从该最终表示进行投影并通过各自层专属的权重得到不同结果但这种做法为什么可行呢他们的理由是这是模型当前对该token已知信息的一个庞大的学习性表示因为在经过20个Transformer层之后每个token表示都已经经历了相当深度的上下文处理一起吧" } ], "id": 268, "review_annotation": "", "row_id": 268, "bounded_row_id": 268, "parent_row_ids": [ 268 ], "parent_row_id": 268, "packet_id": 20, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 550.9597115384616, "packet_end": 580.1476971153846, "raw_text": "并通过各自层专属的权重", "selected_text": "并通过各自层专属的权重", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 268 ], "packet_id": 20, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 564.4997115384616, "end": 565.5797115384617, "text": "得到不同结果", "raw_asr_text": "得到不同结果", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 20, "words": [ { "word": "得", "start": 564.4997115384616, "end": 564.7797115384617, "probability": 0.98388671875 }, { "word": "到", "start": 564.7797115384617, "end": 564.9397115384617, "probability": 0.99951171875 }, { "word": "不同", "start": 564.9397115384617, "end": 565.1197115384616, "probability": 0.99951171875 }, { "word": "结", "start": 565.1197115384616, "end": 565.3397115384616, "probability": 0.98828125 }, { "word": "果", "start": 565.3397115384616, "end": 565.5797115384617, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08723285655283557, "compression_ratio": 1.3487031700288183, "no_speech_prob": 8.208766999473482e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有层都必须共享的同一份KV缓存而是让每个解码器层都对该最终表示进行各自学习到的投影第20层之后的每一层都可以从该最终表示进行投影并通过各自层专属的权重得到不同结果但这种做法为什么可行呢他们的理由是这是模型当前对该token已知信息的一个庞大的学习性表示因为在经过20个Transformer层之后每个token表示都已经经历了相当深度的上下文处理一起吧" } ], "id": 269, "review_annotation": "", "row_id": 269, "bounded_row_id": 269, "parent_row_ids": [ 269 ], "parent_row_id": 269, "packet_id": 20, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 550.9597115384616, "packet_end": 580.1476971153846, "raw_text": "得到不同结果", "selected_text": "得到不同结果", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 269 ], "packet_id": 20, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 565.5797115384617, "end": 567.4797115384616, "text": "但这种做法为什么可行呢", "raw_asr_text": "但这种做法为什么可行呢", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 20, "words": [ { "word": "但", "start": 565.5797115384617, "end": 565.8597115384616, "probability": 0.93310546875 }, { "word": "这", "start": 565.8597115384616, "end": 566.0797115384617, "probability": 0.99755859375 }, { "word": "种", "start": 566.0797115384617, "end": 566.2797115384617, "probability": 0.99658203125 }, { "word": "做", "start": 566.2797115384617, "end": 566.4197115384617, "probability": 0.98974609375 }, { "word": "法", "start": 566.4197115384617, "end": 566.5597115384617, "probability": 1.0 }, { "word": "为什么", "start": 566.5597115384617, "end": 566.8597115384616, "probability": 0.990234375 }, { "word": "可", "start": 566.8597115384616, "end": 567.1197115384616, "probability": 0.99658203125 }, { "word": "行", "start": 567.1197115384616, "end": 567.2797115384617, "probability": 0.99951171875 }, { "word": "呢", "start": 567.2797115384617, "end": 567.4797115384616, "probability": 0.783203125 } ], "asr_metrics": { "avg_logprob": -0.08723285655283557, "compression_ratio": 1.3487031700288183, "no_speech_prob": 8.208766999473482e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有层都必须共享的同一份KV缓存而是让每个解码器层都对该最终表示进行各自学习到的投影第20层之后的每一层都可以从该最终表示进行投影并通过各自层专属的权重得到不同结果但这种做法为什么可行呢他们的理由是这是模型当前对该token已知信息的一个庞大的学习性表示因为在经过20个Transformer层之后每个token表示都已经经历了相当深度的上下文处理一起吧" } ], "id": 270, "review_annotation": "", "row_id": 270, "bounded_row_id": 270, "parent_row_ids": [ 270 ], "parent_row_id": 270, "packet_id": 20, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 550.9597115384616, "packet_end": 580.1476971153846, "raw_text": "但这种做法为什么可行呢", "selected_text": "但这种做法为什么可行呢", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 270 ], "packet_id": 20, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 567.4797115384616, "end": 568.4797115384616, "text": "他们的理由是", "raw_asr_text": "他们的理由是", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 20, "words": [ { "word": "他们", "start": 567.4797115384616, "end": 567.7997115384617, "probability": 0.61181640625 }, { "word": "的", "start": 567.7997115384617, "end": 567.9597115384616, "probability": 0.99951171875 }, { "word": "理", "start": 567.9597115384616, "end": 568.0797115384617, "probability": 1.0 }, { "word": "由", "start": 568.0797115384617, "end": 568.2397115384616, "probability": 1.0 }, { "word": "是", "start": 568.2397115384616, "end": 568.4797115384616, "probability": 0.99365234375 } ], "asr_metrics": { "avg_logprob": -0.08723285655283557, "compression_ratio": 1.3487031700288183, "no_speech_prob": 8.208766999473482e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有层都必须共享的同一份KV缓存而是让每个解码器层都对该最终表示进行各自学习到的投影第20层之后的每一层都可以从该最终表示进行投影并通过各自层专属的权重得到不同结果但这种做法为什么可行呢他们的理由是这是模型当前对该token已知信息的一个庞大的学习性表示因为在经过20个Transformer层之后每个token表示都已经经历了相当深度的上下文处理一起吧" } ], "id": 271, "review_annotation": "", "row_id": 271, "bounded_row_id": 271, "parent_row_ids": [ 271 ], "parent_row_id": 271, "packet_id": 20, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 550.9597115384616, "packet_end": 580.1476971153846, "raw_text": "他们的理由是", "selected_text": "他们的理由是", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 271 ], "packet_id": 20, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 568.4797115384616, "end": 571.8797115384616, "text": "这是模型当前对该token已知信息的", "raw_asr_text": "这是模型当前对该token已知信息的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 20, "words": [ { "word": "这是", "start": 568.4797115384616, "end": 568.7597115384616, "probability": 0.98876953125 }, { "word": "模", "start": 568.7597115384616, "end": 569.0397115384617, "probability": 0.9931640625 }, { "word": "型", "start": 569.0397115384617, "end": 569.2597115384616, "probability": 0.99951171875 }, { "word": "当", "start": 569.2597115384616, "end": 569.4797115384616, "probability": 0.998046875 }, { "word": "前", "start": 569.4797115384616, "end": 569.8597115384616, "probability": 0.99951171875 }, { "word": "对", "start": 569.8597115384616, "end": 570.1797115384617, "probability": 0.99853515625 }, { "word": "该", "start": 570.1797115384617, "end": 570.5797115384617, "probability": 0.99853515625 }, { "word": "t", "start": 570.5797115384617, "end": 570.7397115384616, "probability": 0.58349609375 }, { "word": "oken", "start": 570.7397115384616, "end": 570.9397115384617, "probability": 0.9970703125 }, { "word": "已", "start": 570.9397115384617, "end": 571.2197115384616, "probability": 0.97021484375 }, { "word": "知", "start": 571.2197115384616, "end": 571.3997115384617, "probability": 0.9970703125 }, { "word": "信", "start": 571.3997115384617, "end": 571.5997115384616, "probability": 0.998046875 }, { "word": "息", "start": 571.5997115384616, "end": 571.7797115384617, "probability": 1.0 }, { "word": "的", "start": 571.7797115384617, "end": 571.8797115384616, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.08723285655283557, "compression_ratio": 1.3487031700288183, "no_speech_prob": 8.208766999473482e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有层都必须共享的同一份KV缓存而是让每个解码器层都对该最终表示进行各自学习到的投影第20层之后的每一层都可以从该最终表示进行投影并通过各自层专属的权重得到不同结果但这种做法为什么可行呢他们的理由是这是模型当前对该token已知信息的一个庞大的学习性表示因为在经过20个Transformer层之后每个token表示都已经经历了相当深度的上下文处理一起吧" } ], "id": 272, "review_annotation": "", "row_id": 272, "bounded_row_id": 272, "parent_row_ids": [ 272 ], "parent_row_id": 272, "packet_id": 20, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 550.9597115384616, "packet_end": 580.1476971153846, "raw_text": "这是模型当前对该token已知信息的", "selected_text": "这是模型当前对该token已知信息的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 272 ], "packet_id": 20, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 571.8797115384616, "end": 573.6997115384617, "text": "一个庞大的学习性表示", "raw_asr_text": "一个庞大的学习性表示", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 20, "words": [ { "word": "一个", "start": 571.8797115384616, "end": 572.0397115384617, "probability": 0.96142578125 }, { "word": "庞", "start": 572.0397115384617, "end": 572.4797115384616, "probability": 0.9951171875 }, { "word": "大的", "start": 572.4797115384616, "end": 572.8797115384616, "probability": 0.99951171875 }, { "word": "学", "start": 572.8797115384616, "end": 573.1597115384617, "probability": 0.99951171875 }, { "word": "习", "start": 573.1597115384617, "end": 573.2797115384617, "probability": 1.0 }, { "word": "性", "start": 573.2797115384617, "end": 573.3997115384617, "probability": 0.99658203125 }, { "word": "表示", "start": 573.3997115384617, "end": 573.6997115384617, "probability": 0.99658203125 } ], "asr_metrics": { "avg_logprob": -0.08723285655283557, "compression_ratio": 1.3487031700288183, "no_speech_prob": 8.208766999473482e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有层都必须共享的同一份KV缓存而是让每个解码器层都对该最终表示进行各自学习到的投影第20层之后的每一层都可以从该最终表示进行投影并通过各自层专属的权重得到不同结果但这种做法为什么可行呢他们的理由是这是模型当前对该token已知信息的一个庞大的学习性表示因为在经过20个Transformer层之后每个token表示都已经经历了相当深度的上下文处理一起吧" } ], "id": 273, "review_annotation": "", "row_id": 273, "bounded_row_id": 273, "parent_row_ids": [ 273 ], "parent_row_id": 273, "packet_id": 20, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 550.9597115384616, "packet_end": 580.1476971153846, "raw_text": "一个庞大的学习性表示", "selected_text": "一个庞大的学习性表示", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 273 ], "packet_id": 20, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 573.6997115384617, "end": 576.1197115384616, "text": "因为在经过20个Transformer层之后", "raw_asr_text": "因为在经过20个Transformer层之后", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 20, "words": [ { "word": "因为", "start": 573.6997115384617, "end": 574.2597115384616, "probability": 0.99658203125 }, { "word": "在", "start": 574.2597115384616, "end": 574.4197115384617, "probability": 0.99755859375 }, { "word": "经", "start": 574.4197115384617, "end": 574.6197115384616, "probability": 0.99853515625 }, { "word": "过", "start": 574.6197115384616, "end": 574.7997115384617, "probability": 0.99951171875 }, { "word": "20", "start": 574.7997115384617, "end": 574.9997115384616, "probability": 0.9931640625 }, { "word": "个", "start": 574.9997115384616, "end": 575.1197115384616, "probability": 0.98193359375 }, { "word": "Trans", "start": 575.1197115384616, "end": 575.2597115384616, "probability": 0.55419921875 }, { "word": "form", "start": 575.2597115384616, "end": 575.4597115384616, "probability": 0.99853515625 }, { "word": "er", "start": 575.4597115384616, "end": 575.5597115384617, "probability": 0.9580078125 }, { "word": "层", "start": 575.5597115384617, "end": 575.7197115384616, "probability": 0.99658203125 }, { "word": "之", "start": 575.7197115384616, "end": 575.9197115384617, "probability": 0.986328125 }, { "word": "后", "start": 575.9197115384617, "end": 576.1197115384616, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08723285655283557, "compression_ratio": 1.3487031700288183, "no_speech_prob": 8.208766999473482e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有层都必须共享的同一份KV缓存而是让每个解码器层都对该最终表示进行各自学习到的投影第20层之后的每一层都可以从该最终表示进行投影并通过各自层专属的权重得到不同结果但这种做法为什么可行呢他们的理由是这是模型当前对该token已知信息的一个庞大的学习性表示因为在经过20个Transformer层之后每个token表示都已经经历了相当深度的上下文处理一起吧" } ], "id": 274, "review_annotation": "", "row_id": 274, "bounded_row_id": 274, "parent_row_ids": [ 274 ], "parent_row_id": 274, "packet_id": 20, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 550.9597115384616, "packet_end": 580.1476971153846, "raw_text": "因为在经过20个Transformer层之后", "selected_text": "因为在经过20个Transformer层之后", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 274 ], "packet_id": 20, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 576.1197115384616, "end": 578.0597115384617, "text": "每个token表示都已经经历了", "raw_asr_text": "每个token表示都已经经历了", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 20, "words": [ { "word": "每", "start": 576.1197115384616, "end": 576.4397115384617, "probability": 0.998046875 }, { "word": "个", "start": 576.4397115384617, "end": 576.5797115384617, "probability": 0.99755859375 }, { "word": "t", "start": 576.5797115384617, "end": 576.6997115384617, "probability": 0.818359375 }, { "word": "oken", "start": 576.6997115384617, "end": 576.8397115384616, "probability": 0.9990234375 }, { "word": "表示", "start": 576.8397115384616, "end": 577.0597115384617, "probability": 0.9912109375 }, { "word": "都", "start": 577.0597115384617, "end": 577.3797115384616, "probability": 0.99853515625 }, { "word": "已经", "start": 577.3797115384616, "end": 577.5997115384616, "probability": 0.998046875 }, { "word": "经", "start": 577.5997115384616, "end": 577.7797115384617, "probability": 0.9990234375 }, { "word": "历", "start": 577.7797115384617, "end": 577.9397115384617, "probability": 0.99853515625 }, { "word": "了", "start": 577.9397115384617, "end": 578.0597115384617, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.08723285655283557, "compression_ratio": 1.3487031700288183, "no_speech_prob": 8.208766999473482e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有层都必须共享的同一份KV缓存而是让每个解码器层都对该最终表示进行各自学习到的投影第20层之后的每一层都可以从该最终表示进行投影并通过各自层专属的权重得到不同结果但这种做法为什么可行呢他们的理由是这是模型当前对该token已知信息的一个庞大的学习性表示因为在经过20个Transformer层之后每个token表示都已经经历了相当深度的上下文处理一起吧" } ], "id": 275, "review_annotation": "", "row_id": 275, "bounded_row_id": 275, "parent_row_ids": [ 275 ], "parent_row_id": 275, "packet_id": 20, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 550.9597115384616, "packet_end": 580.1476971153846, "raw_text": "每个token表示都已经经历了", "selected_text": "每个token表示都已经经历了", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 275 ], "packet_id": 20, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 578.0597115384617, "end": 579.3476971153847, "text": "相当深度的上下文处理", "raw_asr_text": "相当深度的上下文处理", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 20, "words": [ { "word": "相", "start": 578.0597115384617, "end": 578.2997115384617, "probability": 0.998046875 }, { "word": "当", "start": 578.2997115384617, "end": 578.5797115384617, "probability": 0.9990234375 }, { "word": "深", "start": 578.5797115384617, "end": 578.8197115384617, "probability": 0.99951171875 }, { "word": "度", "start": 578.8197115384617, "end": 578.9797115384616, "probability": 0.99951171875 }, { "word": "的", "start": 578.9797115384616, "end": 579.0997115384616, "probability": 0.99853515625 }, { "word": "上", "start": 579.0997115384616, "end": 579.2397115384616, "probability": 0.99853515625 }, { "word": "下", "start": 579.2397115384616, "end": 579.3997115384617, "probability": 0.9990234375 }, { "word": "文", "start": 579.3997115384617, "end": 579.5397115384617, "probability": 0.97509765625 }, { "word": "处", "start": 579.5397115384617, "end": 579.6997115384617, "probability": 0.9990234375 }, { "word": "理", "start": 579.6997115384617, "end": 579.8197115384617, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08723285655283557, "compression_ratio": 1.3487031700288183, "no_speech_prob": 8.208766999473482e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所有层都必须共享的同一份KV缓存而是让每个解码器层都对该最终表示进行各自学习到的投影第20层之后的每一层都可以从该最终表示进行投影并通过各自层专属的权重得到不同结果但这种做法为什么可行呢他们的理由是这是模型当前对该token已知信息的一个庞大的学习性表示因为在经过20个Transformer层之后每个token表示都已经经历了相当深度的上下文处理一起吧" } ], "id": 276, "review_annotation": "", "row_id": 276, "bounded_row_id": 276, "parent_row_ids": [ 276 ], "parent_row_id": 276, "packet_id": 20, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 550.9597115384616, "packet_end": 580.1476971153846, "raw_text": "相当深度的上下文处理", "selected_text": "相当深度的上下文处理", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 276 ], "packet_id": 20, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 579.8076971153847, "end": 580.9076971153847, "text": "基本上已经成为一个", "raw_asr_text": "基本上已经成为一个", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "基本", "start": 579.8076971153847, "end": 580.1276971153848, "probability": 0.9775390625 }, { "word": "上", "start": 580.1276971153848, "end": 580.3076971153847, "probability": 0.99658203125 }, { "word": "已经", "start": 580.3076971153847, "end": 580.4676971153847, "probability": 0.99609375 }, { "word": "成", "start": 580.4676971153847, "end": 580.6476971153847, "probability": 0.9990234375 }, { "word": "为", "start": 580.6476971153847, "end": 580.7676971153847, "probability": 0.9951171875 }, { "word": "一个", "start": 580.7676971153847, "end": 580.9076971153847, "probability": 0.990234375 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 277, "review_annotation": "", "row_id": 277, "bounded_row_id": 277, "parent_row_ids": [ 277 ], "parent_row_id": 277, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "基本上已经成为一个", "selected_text": "基本上已经成为一个", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 277 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 580.9076971153847, "end": 582.3076971153847, "text": "丰富的公共信息源", "raw_asr_text": "丰富的公共信息源", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "丰", "start": 580.9076971153847, "end": 581.1876971153847, "probability": 0.998046875 }, { "word": "富", "start": 581.1876971153847, "end": 581.3076971153847, "probability": 0.99951171875 }, { "word": "的", "start": 581.3076971153847, "end": 581.4476971153847, "probability": 0.99951171875 }, { "word": "公", "start": 581.4476971153847, "end": 581.6276971153848, "probability": 0.99951171875 }, { "word": "共", "start": 581.6276971153848, "end": 581.7876971153847, "probability": 0.99462890625 }, { "word": "信", "start": 581.7876971153847, "end": 581.9476971153847, "probability": 0.99951171875 }, { "word": "息", "start": 581.9476971153847, "end": 582.1076971153847, "probability": 0.99853515625 }, { "word": "源", "start": 582.1076971153847, "end": 582.3076971153847, "probability": 0.98046875 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 278, "review_annotation": "", "row_id": 278, "bounded_row_id": 278, "parent_row_ids": [ 278 ], "parent_row_id": 278, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "丰富的公共信息源", "selected_text": "丰富的公共信息源", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 278 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 582.3076971153847, "end": 584.2076971153847, "text": "注意力所使用的K和V", "raw_asr_text": "注意力所使用的K和V", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "注意", "start": 582.3076971153847, "end": 582.6476971153847, "probability": 0.939453125 }, { "word": "力", "start": 582.6476971153847, "end": 582.8076971153847, "probability": 0.99853515625 }, { "word": "所", "start": 582.8076971153847, "end": 582.9876971153848, "probability": 0.99755859375 }, { "word": "使", "start": 582.9876971153848, "end": 583.2076971153847, "probability": 0.99853515625 }, { "word": "用", "start": 583.2076971153847, "end": 583.3276971153847, "probability": 1.0 }, { "word": "的", "start": 583.3276971153847, "end": 583.4476971153847, "probability": 0.99853515625 }, { "word": "K", "start": 583.4476971153847, "end": 583.6676971153847, "probability": 0.84033203125 }, { "word": "和", "start": 583.6676971153847, "end": 583.9676971153847, "probability": 0.94482421875 }, { "word": "V", "start": 583.9676971153847, "end": 584.2076971153847, "probability": 0.9912109375 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 279, "review_annotation": "", "row_id": 279, "bounded_row_id": 279, "parent_row_ids": [ 279 ], "parent_row_id": 279, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "注意力所使用的K和V", "selected_text": "注意力所使用的K和V", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 279 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 584.2076971153847, "end": 584.9676971153847, "text": "基本上就是", "raw_asr_text": "基本上就是", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "基本", "start": 584.2076971153847, "end": 584.5876971153847, "probability": 0.98876953125 }, { "word": "上", "start": 584.5876971153847, "end": 584.7676971153847, "probability": 0.99755859375 }, { "word": "就是", "start": 584.7676971153847, "end": 584.9676971153847, "probability": 0.9912109375 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 280, "review_annotation": "", "row_id": 280, "bounded_row_id": 280, "parent_row_ids": [ 280 ], "parent_row_id": 280, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "基本上就是", "selected_text": "基本上就是", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 280 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 584.9676971153847, "end": 586.8476971153847, "text": "该信息源的不同学习试图", "raw_asr_text": "该信息源的不同学习试图", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "该", "start": 584.9676971153847, "end": 585.2276971153847, "probability": 0.9951171875 }, { "word": "信", "start": 585.2276971153847, "end": 585.4076971153847, "probability": 0.99951171875 }, { "word": "息", "start": 585.4076971153847, "end": 585.5876971153847, "probability": 1.0 }, { "word": "源", "start": 585.5876971153847, "end": 585.6876971153847, "probability": 0.9990234375 }, { "word": "的", "start": 585.6876971153847, "end": 585.8276971153847, "probability": 0.99951171875 }, { "word": "不同", "start": 585.8276971153847, "end": 586.1076971153847, "probability": 0.99755859375 }, { "word": "学", "start": 586.1076971153847, "end": 586.3476971153847, "probability": 0.99853515625 }, { "word": "习", "start": 586.3476971153847, "end": 586.4876971153848, "probability": 0.9990234375 }, { "word": "试", "start": 586.4876971153848, "end": 586.6276971153848, "probability": 0.6884765625 }, { "word": "图", "start": 586.6276971153848, "end": 586.8476971153847, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 281, "review_annotation": "", "row_id": 281, "bounded_row_id": 281, "parent_row_ids": [ 281 ], "parent_row_id": 281, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "该信息源的不同学习试图", "selected_text": "该信息源的不同学习试图", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 281 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 586.8476971153847, "end": 588.4676971153847, "text": "建会确定这条信息", "raw_asr_text": "建会确定这条信息", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "建", "start": 586.8476971153847, "end": 587.2076971153847, "probability": 0.1671142578125 }, { "word": "会", "start": 587.2076971153847, "end": 587.4276971153847, "probability": 0.8310546875 }, { "word": "确", "start": 587.4276971153847, "end": 587.6476971153847, "probability": 0.9990234375 }, { "word": "定", "start": 587.6476971153847, "end": 587.7876971153847, "probability": 0.99755859375 }, { "word": "这", "start": 587.7876971153847, "end": 587.9476971153847, "probability": 0.9970703125 }, { "word": "条", "start": 587.9476971153847, "end": 588.0676971153847, "probability": 0.9990234375 }, { "word": "信", "start": 588.0676971153847, "end": 588.2676971153847, "probability": 0.9990234375 }, { "word": "息", "start": 588.2676971153847, "end": 588.4676971153847, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 282, "review_annotation": "", "row_id": 282, "bounded_row_id": 282, "parent_row_ids": [ 282 ], "parent_row_id": 282, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "建会确定这条信息", "selected_text": "建会确定这条信息", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 282 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 588.4676971153847, "end": 589.2876971153847, "text": "要如何表示", "raw_asr_text": "要如何表示", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "要", "start": 588.4676971153847, "end": 588.6676971153847, "probability": 0.99755859375 }, { "word": "如何", "start": 588.6676971153847, "end": 588.8676971153848, "probability": 0.99951171875 }, { "word": "表示", "start": 588.8676971153848, "end": 589.2876971153847, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 283, "review_annotation": "", "row_id": 283, "bounded_row_id": 283, "parent_row_ids": [ 283 ], "parent_row_id": 283, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "要如何表示", "selected_text": "要如何表示", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 283 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 589.2876971153847, "end": 590.4876971153848, "text": "以便后续的查询", "raw_asr_text": "以便后续的查询", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "以", "start": 589.2876971153847, "end": 589.5676971153847, "probability": 0.99853515625 }, { "word": "便", "start": 589.5676971153847, "end": 589.7076971153847, "probability": 0.97900390625 }, { "word": "后", "start": 589.7076971153847, "end": 589.9076971153847, "probability": 0.998046875 }, { "word": "续", "start": 589.9076971153847, "end": 590.0676971153847, "probability": 0.99951171875 }, { "word": "的", "start": 590.0676971153847, "end": 590.1476971153847, "probability": 0.98876953125 }, { "word": "查", "start": 590.1476971153847, "end": 590.2876971153847, "probability": 0.9951171875 }, { "word": "询", "start": 590.2876971153847, "end": 590.4876971153848, "probability": 0.99365234375 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 284, "review_annotation": "", "row_id": 284, "bounded_row_id": 284, "parent_row_ids": [ 284 ], "parent_row_id": 284, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "以便后续的查询", "selected_text": "以便后续的查询", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 284 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 590.4876971153848, "end": 591.4876971153848, "text": "能够与它匹配", "raw_asr_text": "能够与它匹配", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "能", "start": 590.4876971153848, "end": 590.6476971153847, "probability": 0.998046875 }, { "word": "够", "start": 590.6476971153847, "end": 590.8476971153847, "probability": 1.0 }, { "word": "与", "start": 590.8476971153847, "end": 590.9476971153847, "probability": 1.0 }, { "word": "它", "start": 590.9476971153847, "end": 591.0876971153847, "probability": 0.88818359375 }, { "word": "匹", "start": 591.0876971153847, "end": 591.3076971153847, "probability": 0.9951171875 }, { "word": "配", "start": 591.3076971153847, "end": 591.4876971153848, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 285, "review_annotation": "", "row_id": 285, "bounded_row_id": 285, "parent_row_ids": [ 285 ], "parent_row_id": 285, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "能够与它匹配", "selected_text": "能够与它匹配", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 285 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 591.8476971153847, "end": 593.9876971153848, "text": "值则决定当它被注意到时", "raw_asr_text": "值则决定当它被注意到时", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "值", "start": 591.8476971153847, "end": 592.2076971153847, "probability": 0.9072265625 }, { "word": "则", "start": 592.2076971153847, "end": 592.4876971153848, "probability": 0.994140625 }, { "word": "决", "start": 592.4876971153848, "end": 592.6476971153847, "probability": 0.9990234375 }, { "word": "定", "start": 592.6476971153847, "end": 592.7676971153847, "probability": 1.0 }, { "word": "当", "start": 592.7676971153847, "end": 592.9476971153847, "probability": 0.99609375 }, { "word": "它", "start": 592.9476971153847, "end": 593.1476971153847, "probability": 0.88623046875 }, { "word": "被", "start": 593.1476971153847, "end": 593.3076971153847, "probability": 0.9990234375 }, { "word": "注意", "start": 593.3076971153847, "end": 593.5276971153847, "probability": 0.99853515625 }, { "word": "到", "start": 593.5276971153847, "end": 593.7476971153848, "probability": 0.998046875 }, { "word": "时", "start": 593.7476971153848, "end": 593.9876971153848, "probability": 0.99560546875 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 286, "review_annotation": "", "row_id": 286, "bounded_row_id": 286, "parent_row_ids": [ 286 ], "parent_row_id": 286, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "值则决定当它被注意到时", "selected_text": "值则决定当它被注意到时", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 286 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 593.9876971153848, "end": 595.5276971153847, "text": "应当返回哪一项内容", "raw_asr_text": "应当返回哪一项内容", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "应", "start": 593.9876971153848, "end": 594.2476971153848, "probability": 0.99609375 }, { "word": "当", "start": 594.2476971153848, "end": 594.4076971153847, "probability": 0.99169921875 }, { "word": "返", "start": 594.4076971153847, "end": 594.6676971153847, "probability": 0.9951171875 }, { "word": "回", "start": 594.6676971153847, "end": 594.8276971153847, "probability": 0.99951171875 }, { "word": "哪", "start": 594.8276971153847, "end": 595.0076971153848, "probability": 0.9814453125 }, { "word": "一", "start": 595.0076971153848, "end": 595.1276971153848, "probability": 0.994140625 }, { "word": "项", "start": 595.1276971153848, "end": 595.2476971153848, "probability": 0.9990234375 }, { "word": "内", "start": 595.2476971153848, "end": 595.3876971153848, "probability": 0.9990234375 }, { "word": "容", "start": 595.3876971153848, "end": 595.5276971153847, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 287, "review_annotation": "", "row_id": 287, "bounded_row_id": 287, "parent_row_ids": [ 287 ], "parent_row_id": 287, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "应当返回哪一项内容", "selected_text": "应当返回哪一项内容", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 287 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 595.5276971153847, "end": 596.5476971153847, "text": "所以通常情况下", "raw_asr_text": "所以通常情况下", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "所以", "start": 595.5276971153847, "end": 595.7676971153847, "probability": 0.98974609375 }, { "word": "通", "start": 595.7676971153847, "end": 595.9476971153847, "probability": 0.998046875 }, { "word": "常", "start": 595.9476971153847, "end": 596.1476971153847, "probability": 1.0 }, { "word": "情况", "start": 596.1476971153847, "end": 596.3476971153847, "probability": 0.9970703125 }, { "word": "下", "start": 596.3476971153847, "end": 596.5476971153847, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 288, "review_annotation": "", "row_id": 288, "bounded_row_id": 288, "parent_row_ids": [ 288 ], "parent_row_id": 288, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "所以通常情况下", "selected_text": "所以通常情况下", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 288 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 596.5476971153847, "end": 597.3676971153848, "text": "如果你想得到", "raw_asr_text": "如果你想得到", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "如果", "start": 596.5476971153847, "end": 596.7676971153847, "probability": 0.994140625 }, { "word": "你想", "start": 596.7676971153847, "end": 597.0276971153847, "probability": 0.99365234375 }, { "word": "得", "start": 597.0276971153847, "end": 597.2076971153847, "probability": 0.9990234375 }, { "word": "到", "start": 597.2076971153847, "end": 597.3676971153848, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 289, "review_annotation": "", "row_id": 289, "bounded_row_id": 289, "parent_row_ids": [ 289 ], "parent_row_id": 289, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "如果你想得到", "selected_text": "如果你想得到", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 289 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 597.3676971153848, "end": 598.7676971153847, "text": "第35层的KV", "raw_asr_text": "第35层的KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "第", "start": 597.3676971153848, "end": 597.5476971153847, "probability": 0.96484375 }, { "word": "35", "start": 597.5476971153847, "end": 597.9676971153847, "probability": 0.99658203125 }, { "word": "层", "start": 597.9676971153847, "end": 598.2676971153847, "probability": 0.99951171875 }, { "word": "的", "start": 598.2676971153847, "end": 598.4076971153847, "probability": 0.99755859375 }, { "word": "K", "start": 598.4076971153847, "end": 598.5676971153847, "probability": 0.9638671875 }, { "word": "V", "start": 598.5676971153847, "end": 598.7676971153847, "probability": 0.9638671875 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 290, "review_annotation": "", "row_id": 290, "bounded_row_id": 290, "parent_row_ids": [ 290 ], "parent_row_id": 290, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "第35层的KV", "selected_text": "第35层的KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 290 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 598.7676971153847, "end": 599.9676971153847, "text": "每个历史提示词", "raw_asr_text": "每个历史提示词", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "每", "start": 598.7676971153847, "end": 599.1276971153848, "probability": 0.99365234375 }, { "word": "个", "start": 599.1276971153848, "end": 599.2676971153847, "probability": 0.99755859375 }, { "word": "历", "start": 599.2676971153847, "end": 599.4076971153847, "probability": 0.998046875 }, { "word": "史", "start": 599.4076971153847, "end": 599.5476971153847, "probability": 1.0 }, { "word": "提", "start": 599.5476971153847, "end": 599.6676971153847, "probability": 0.99853515625 }, { "word": "示", "start": 599.6676971153847, "end": 599.8076971153847, "probability": 0.9990234375 }, { "word": "词", "start": 599.8076971153847, "end": 599.9676971153847, "probability": 0.97802734375 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 291, "review_annotation": "", "row_id": 291, "bounded_row_id": 291, "parent_row_ids": [ 291 ], "parent_row_id": 291, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "每个历史提示词", "selected_text": "每个历史提示词", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 291 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 599.9676971153847, "end": 600.9076971153847, "text": "Token都必须先", "raw_asr_text": "Token都必须先", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "T", "start": 599.9676971153847, "end": 600.1276971153848, "probability": 0.76318359375 }, { "word": "oken", "start": 600.1276971153848, "end": 600.3276971153847, "probability": 0.98046875 }, { "word": "都", "start": 600.3276971153847, "end": 600.4876971153848, "probability": 0.97998046875 }, { "word": "必", "start": 600.4876971153848, "end": 600.6476971153847, "probability": 0.998046875 }, { "word": "须", "start": 600.6476971153847, "end": 600.7476971153848, "probability": 1.0 }, { "word": "先", "start": 600.7476971153848, "end": 600.9076971153847, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 292, "review_annotation": "", "row_id": 292, "bounded_row_id": 292, "parent_row_ids": [ 292 ], "parent_row_id": 292, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "Token都必须先", "selected_text": "Token都必须先", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 292 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 600.9076971153847, "end": 602.5276971153847, "text": "实际留进第21层", "raw_asr_text": "实际留进第21层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "实", "start": 600.9076971153847, "end": 601.2076971153847, "probability": 0.998046875 }, { "word": "际", "start": 601.2076971153847, "end": 601.3876971153848, "probability": 0.99951171875 }, { "word": "留", "start": 601.3876971153848, "end": 601.5476971153847, "probability": 0.81494140625 }, { "word": "进", "start": 601.5476971153847, "end": 601.7076971153847, "probability": 0.6396484375 }, { "word": "第", "start": 601.7076971153847, "end": 601.8876971153848, "probability": 0.99658203125 }, { "word": "21", "start": 601.8876971153848, "end": 602.1676971153847, "probability": 0.99951171875 }, { "word": "层", "start": 602.1676971153847, "end": 602.5276971153847, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 293, "review_annotation": "", "row_id": 293, "bounded_row_id": 293, "parent_row_ids": [ 293 ], "parent_row_id": 293, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "实际留进第21层", "selected_text": "实际留进第21层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 293 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 602.5276971153847, "end": 603.8276971153847, "text": "一直到第35层", "raw_asr_text": "一直到第35层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "一直", "start": 602.5276971153847, "end": 602.7676971153847, "probability": 0.97021484375 }, { "word": "到", "start": 602.7676971153847, "end": 602.9676971153847, "probability": 0.99951171875 }, { "word": "第", "start": 602.9676971153847, "end": 603.1276971153848, "probability": 0.99609375 }, { "word": "35", "start": 603.1276971153848, "end": 603.4676971153847, "probability": 0.99853515625 }, { "word": "层", "start": 603.4676971153847, "end": 603.8276971153847, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 294, "review_annotation": "", "row_id": 294, "bounded_row_id": 294, "parent_row_ids": [ 294 ], "parent_row_id": 294, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "一直到第35层", "selected_text": "一直到第35层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 294 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 603.8276971153847, "end": 605.1476971153847, "text": "然后第35层", "raw_asr_text": "然后第35层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "然后", "start": 603.8276971153847, "end": 604.1476971153847, "probability": 0.99169921875 }, { "word": "第", "start": 604.1476971153847, "end": 604.4076971153847, "probability": 0.99560546875 }, { "word": "35", "start": 604.4076971153847, "end": 604.7676971153847, "probability": 0.9990234375 }, { "word": "层", "start": 604.7676971153847, "end": 605.1476971153847, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 295, "review_annotation": "", "row_id": 295, "bounded_row_id": 295, "parent_row_ids": [ 295 ], "parent_row_id": 295, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "然后第35层", "selected_text": "然后第35层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 295 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 605.1476971153847, "end": 606.7676971153847, "text": "再把新的隐藏状态", "raw_asr_text": "再把新的隐藏状态", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 21, "words": [ { "word": "再", "start": 605.1476971153847, "end": 605.3476971153847, "probability": 0.990234375 }, { "word": "把", "start": 605.3476971153847, "end": 605.5076971153848, "probability": 0.99951171875 }, { "word": "新", "start": 605.5076971153848, "end": 605.7476971153848, "probability": 0.99951171875 }, { "word": "的", "start": 605.7476971153848, "end": 605.9276971153847, "probability": 0.99951171875 }, { "word": "隐", "start": 605.9276971153847, "end": 606.1876971153847, "probability": 0.998046875 }, { "word": "藏", "start": 606.1876971153847, "end": 606.3476971153847, "probability": 1.0 }, { "word": "状", "start": 606.3476971153847, "end": 606.5476971153847, "probability": 0.998046875 }, { "word": "态", "start": 606.5476971153847, "end": 606.7676971153847, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09911909015900498, "compression_ratio": 1.3100558659217878, "no_speech_prob": 6.811318870036942e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "深度的上下文处理基本上已经成为一个丰富的公共信息源注意力所使用的K和V基本上就是该信息源的不同学习试图建会确定这条信息要如何表示以便后续的查询能够与它匹配值则决定当它被注意到时应当返回哪一项内容所以通常情况下如果你想得到第35层的KV每个历史提示词Token都必须先实际留进第21层一直到第35层然后第35层再把新的隐藏状态投影成它的KV" } ], "id": 296, "review_annotation": "", "row_id": 296, "bounded_row_id": 296, "parent_row_ids": [ 296 ], "parent_row_id": 296, "packet_id": 21, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 578.5476971153847, "packet_end": 607.7356826923077, "raw_text": "再把新的隐藏状态", "selected_text": "再把新的隐藏状态", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 296 ], "packet_id": 21, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 606.9356826923078, "end": 607.8356826923078, "text": "投影成它的KV", "raw_asr_text": "投影成它的KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 22, "words": [ { "word": "投", "start": 606.5756826923079, "end": 606.9356826923078, "probability": 0.0489501953125 }, { "word": "影", "start": 606.9356826923078, "end": 607.1156826923078, "probability": 0.99072265625 }, { "word": "成", "start": 607.1156826923078, "end": 607.2756826923078, "probability": 0.9970703125 }, { "word": "它的", "start": 607.2756826923078, "end": 607.5156826923078, "probability": 0.75830078125 }, { "word": "K", "start": 607.5156826923078, "end": 607.6556826923078, "probability": 0.8701171875 }, { "word": "V", "start": 607.6556826923078, "end": 607.8356826923078, "probability": 0.99365234375 } ], "asr_metrics": { "avg_logprob": -0.09344960772801959, "compression_ratio": 1.206798866855524, "no_speech_prob": 7.747375657674027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投影成它的KV这里DeepSeek基本上是在质疑仅仅为了创建全局记忆库是否真的需要所有这些非线性计算于是直接砍掉中间的每一个串行流程使用该隐藏状态再交给第35层投影借此获得KV信息不过如果你还记得前面讲过的内容DeepSeek其实并不是每次都为每一层生成完全独立的庞大KV缓存CSAR已经通过三种不同模式让相邻层共享全局KV所以只有" } ], "id": 297, "review_annotation": "", "row_id": 297, "bounded_row_id": 297, "parent_row_ids": [ 297 ], "parent_row_id": 297, "packet_id": 22, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 606.1356826923078, "packet_end": 635.3236682692308, "raw_text": "投影成它的KV", "selected_text": "投影成它的KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 297 ], "packet_id": 22, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 607.8356826923078, "end": 609.8556826923078, "text": "这里DeepSeek基本上是在质疑", "raw_asr_text": "这里DeepSeek基本上是在质疑", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 22, "words": [ { "word": "这里", "start": 607.8356826923078, "end": 608.1756826923078, "probability": 0.97314453125 }, { "word": "De", "start": 608.1756826923078, "end": 608.3356826923078, "probability": 0.89453125 }, { "word": "ep", "start": 608.3356826923078, "end": 608.4356826923078, "probability": 0.970703125 }, { "word": "Se", "start": 608.4356826923078, "end": 608.5756826923079, "probability": 0.419921875 }, { "word": "ek", "start": 608.5756826923079, "end": 608.6756826923078, "probability": 0.75244140625 }, { "word": "基本", "start": 608.6756826923078, "end": 608.9756826923078, "probability": 0.99462890625 }, { "word": "上", "start": 608.9756826923078, "end": 609.1956826923077, "probability": 0.99658203125 }, { "word": "是", "start": 609.1956826923077, "end": 609.3156826923077, "probability": 0.99755859375 }, { "word": "在", "start": 609.3156826923077, "end": 609.4956826923078, "probability": 0.99658203125 }, { "word": "质", "start": 609.4956826923078, "end": 609.6956826923077, "probability": 0.99755859375 }, { "word": "疑", "start": 609.6956826923077, "end": 609.8556826923078, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09344960772801959, "compression_ratio": 1.206798866855524, "no_speech_prob": 7.747375657674027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投影成它的KV这里DeepSeek基本上是在质疑仅仅为了创建全局记忆库是否真的需要所有这些非线性计算于是直接砍掉中间的每一个串行流程使用该隐藏状态再交给第35层投影借此获得KV信息不过如果你还记得前面讲过的内容DeepSeek其实并不是每次都为每一层生成完全独立的庞大KV缓存CSAR已经通过三种不同模式让相邻层共享全局KV所以只有" } ], "id": 298, "review_annotation": "", "row_id": 298, "bounded_row_id": 298, "parent_row_ids": [ 298 ], "parent_row_id": 298, "packet_id": 22, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 606.1356826923078, "packet_end": 635.3236682692308, "raw_text": "这里DeepSeek基本上是在质疑", "selected_text": "这里DeepSeek基本上是在质疑", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 298 ], "packet_id": 22, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 609.8556826923078, "end": 611.9156826923078, "text": "仅仅为了创建全局记忆库", "raw_asr_text": "仅仅为了创建全局记忆库", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 22, "words": [ { "word": "仅", "start": 609.8556826923078, "end": 610.1956826923077, "probability": 0.994140625 }, { "word": "仅", "start": 610.1956826923077, "end": 610.3756826923078, "probability": 0.9990234375 }, { "word": "为", "start": 610.3756826923078, "end": 610.5156826923078, "probability": 0.99951171875 }, { "word": "了", "start": 610.5156826923078, "end": 610.6556826923078, "probability": 0.99951171875 }, { "word": "创", "start": 610.6556826923078, "end": 610.8356826923078, "probability": 1.0 }, { "word": "建", "start": 610.8356826923078, "end": 611.0156826923078, "probability": 1.0 }, { "word": "全", "start": 611.0156826923078, "end": 611.3156826923077, "probability": 0.994140625 }, { "word": "局", "start": 611.3156826923077, "end": 611.4756826923078, "probability": 0.99755859375 }, { "word": "记", "start": 611.4756826923078, "end": 611.6556826923078, "probability": 0.94921875 }, { "word": "忆", "start": 611.6556826923078, "end": 611.7556826923078, "probability": 0.9990234375 }, { "word": "库", "start": 611.7556826923078, "end": 611.9156826923078, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09344960772801959, "compression_ratio": 1.206798866855524, "no_speech_prob": 7.747375657674027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投影成它的KV这里DeepSeek基本上是在质疑仅仅为了创建全局记忆库是否真的需要所有这些非线性计算于是直接砍掉中间的每一个串行流程使用该隐藏状态再交给第35层投影借此获得KV信息不过如果你还记得前面讲过的内容DeepSeek其实并不是每次都为每一层生成完全独立的庞大KV缓存CSAR已经通过三种不同模式让相邻层共享全局KV所以只有" } ], "id": 299, "review_annotation": "", "row_id": 299, "bounded_row_id": 299, "parent_row_ids": [ 299 ], "parent_row_id": 299, "packet_id": 22, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 606.1356826923078, "packet_end": 635.3236682692308, "raw_text": "仅仅为了创建全局记忆库", "selected_text": "仅仅为了创建全局记忆库", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 299 ], "packet_id": 22, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 611.9156826923078, "end": 614.4956826923078, "text": "是否真的需要所有这些非线性计算", "raw_asr_text": "是否真的需要所有这些非线性计算", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 22, "words": [ { "word": "是", "start": 611.9156826923078, "end": 612.2356826923078, "probability": 0.9970703125 }, { "word": "否", "start": 612.2356826923078, "end": 612.3556826923078, "probability": 1.0 }, { "word": "真的", "start": 612.3556826923078, "end": 612.6556826923078, "probability": 0.9990234375 }, { "word": "需要", "start": 612.6556826923078, "end": 612.9956826923078, "probability": 0.99951171875 }, { "word": "所有", "start": 612.9956826923078, "end": 613.4556826923078, "probability": 0.9990234375 }, { "word": "这", "start": 613.4556826923078, "end": 613.6356826923078, "probability": 0.99658203125 }, { "word": "些", "start": 613.6356826923078, "end": 613.7756826923078, "probability": 1.0 }, { "word": "非", "start": 613.7756826923078, "end": 613.9356826923078, "probability": 0.97607421875 }, { "word": "线", "start": 613.9356826923078, "end": 614.0756826923079, "probability": 0.99462890625 }, { "word": "性", "start": 614.0756826923079, "end": 614.1956826923077, "probability": 0.93798828125 }, { "word": "计", "start": 614.1956826923077, "end": 614.3556826923078, "probability": 0.9990234375 }, { "word": "算", "start": 614.3556826923078, "end": 614.4956826923078, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09344960772801959, "compression_ratio": 1.206798866855524, "no_speech_prob": 7.747375657674027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投影成它的KV这里DeepSeek基本上是在质疑仅仅为了创建全局记忆库是否真的需要所有这些非线性计算于是直接砍掉中间的每一个串行流程使用该隐藏状态再交给第35层投影借此获得KV信息不过如果你还记得前面讲过的内容DeepSeek其实并不是每次都为每一层生成完全独立的庞大KV缓存CSAR已经通过三种不同模式让相邻层共享全局KV所以只有" } ], "id": 300, "review_annotation": "", "row_id": 300, "bounded_row_id": 300, "parent_row_ids": [ 300 ], "parent_row_id": 300, "packet_id": 22, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 606.1356826923078, "packet_end": 635.3236682692308, "raw_text": "是否真的需要所有这些非线性计算", "selected_text": "是否真的需要所有这些非线性计算", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 300 ], "packet_id": 22, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 614.4956826923078, "end": 617.5756826923079, "text": "于是直接砍掉中间的每一个串行流程", "raw_asr_text": "于是直接砍掉中间的每一个串行流程", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 22, "words": [ { "word": "于", "start": 614.4956826923078, "end": 614.9156826923078, "probability": 0.9931640625 }, { "word": "是", "start": 614.9156826923078, "end": 615.0556826923078, "probability": 1.0 }, { "word": "直接", "start": 615.0556826923078, "end": 615.3956826923078, "probability": 0.99951171875 }, { "word": "砍", "start": 615.3956826923078, "end": 615.6156826923078, "probability": 0.9970703125 }, { "word": "掉", "start": 615.6156826923078, "end": 615.7556826923078, "probability": 0.99951171875 }, { "word": "中", "start": 615.7556826923078, "end": 615.9156826923078, "probability": 0.9990234375 }, { "word": "间", "start": 615.9156826923078, "end": 616.0556826923078, "probability": 0.99951171875 }, { "word": "的", "start": 616.0556826923078, "end": 616.1756826923078, "probability": 1.0 }, { "word": "每", "start": 616.1756826923078, "end": 616.4756826923078, "probability": 0.99951171875 }, { "word": "一个", "start": 616.4756826923078, "end": 616.8156826923077, "probability": 0.99853515625 }, { "word": "串", "start": 616.8156826923077, "end": 617.0356826923078, "probability": 0.98828125 }, { "word": "行", "start": 617.0356826923078, "end": 617.1956826923077, "probability": 0.45556640625 }, { "word": "流", "start": 617.1956826923077, "end": 617.3556826923078, "probability": 0.99853515625 }, { "word": "程", "start": 617.3556826923078, "end": 617.5756826923079, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09344960772801959, "compression_ratio": 1.206798866855524, "no_speech_prob": 7.747375657674027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投影成它的KV这里DeepSeek基本上是在质疑仅仅为了创建全局记忆库是否真的需要所有这些非线性计算于是直接砍掉中间的每一个串行流程使用该隐藏状态再交给第35层投影借此获得KV信息不过如果你还记得前面讲过的内容DeepSeek其实并不是每次都为每一层生成完全独立的庞大KV缓存CSAR已经通过三种不同模式让相邻层共享全局KV所以只有" } ], "id": 301, "review_annotation": "", "row_id": 301, "bounded_row_id": 301, "parent_row_ids": [ 301 ], "parent_row_id": 301, "packet_id": 22, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 606.1356826923078, "packet_end": 635.3236682692308, "raw_text": "于是直接砍掉中间的每一个串行流程", "selected_text": "于是直接砍掉中间的每一个串行流程", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 301 ], "packet_id": 22, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 617.5756826923079, "end": 618.9956826923078, "text": "使用该隐藏状态", "raw_asr_text": "使用该隐藏状态", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 22, "words": [ { "word": "使", "start": 617.5756826923079, "end": 617.8556826923078, "probability": 0.99658203125 }, { "word": "用", "start": 617.8556826923078, "end": 618.0156826923078, "probability": 1.0 }, { "word": "该", "start": 618.0156826923078, "end": 618.2356826923078, "probability": 0.99951171875 }, { "word": "隐", "start": 618.2356826923078, "end": 618.4556826923078, "probability": 0.9990234375 }, { "word": "藏", "start": 618.4556826923078, "end": 618.6356826923078, "probability": 0.99951171875 }, { "word": "状", "start": 618.6356826923078, "end": 618.8356826923078, "probability": 0.9990234375 }, { "word": "态", "start": 618.8356826923078, "end": 618.9956826923078, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.09344960772801959, "compression_ratio": 1.206798866855524, "no_speech_prob": 7.747375657674027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投影成它的KV这里DeepSeek基本上是在质疑仅仅为了创建全局记忆库是否真的需要所有这些非线性计算于是直接砍掉中间的每一个串行流程使用该隐藏状态再交给第35层投影借此获得KV信息不过如果你还记得前面讲过的内容DeepSeek其实并不是每次都为每一层生成完全独立的庞大KV缓存CSAR已经通过三种不同模式让相邻层共享全局KV所以只有" } ], "id": 302, "review_annotation": "", "row_id": 302, "bounded_row_id": 302, "parent_row_ids": [ 302 ], "parent_row_id": 302, "packet_id": 22, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 606.1356826923078, "packet_end": 635.3236682692308, "raw_text": "使用该隐藏状态", "selected_text": "使用该隐藏状态", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 302 ], "packet_id": 22, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 618.9956826923078, "end": 621.0156826923078, "text": "再交给第35层投影", "raw_asr_text": "再交给第35层投影", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 22, "words": [ { "word": "再", "start": 618.9956826923078, "end": 619.2356826923078, "probability": 0.9931640625 }, { "word": "交", "start": 619.2356826923078, "end": 619.4356826923078, "probability": 0.99169921875 }, { "word": "给", "start": 619.4356826923078, "end": 619.6356826923078, "probability": 0.99951171875 }, { "word": "第", "start": 619.6356826923078, "end": 619.7956826923078, "probability": 0.89501953125 }, { "word": "35", "start": 619.7956826923078, "end": 620.3356826923078, "probability": 0.99853515625 }, { "word": "层", "start": 620.3356826923078, "end": 620.6756826923078, "probability": 0.99755859375 }, { "word": "投", "start": 620.6756826923078, "end": 620.8356826923078, "probability": 1.0 }, { "word": "影", "start": 620.8356826923078, "end": 621.0156826923078, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09344960772801959, "compression_ratio": 1.206798866855524, "no_speech_prob": 7.747375657674027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投影成它的KV这里DeepSeek基本上是在质疑仅仅为了创建全局记忆库是否真的需要所有这些非线性计算于是直接砍掉中间的每一个串行流程使用该隐藏状态再交给第35层投影借此获得KV信息不过如果你还记得前面讲过的内容DeepSeek其实并不是每次都为每一层生成完全独立的庞大KV缓存CSAR已经通过三种不同模式让相邻层共享全局KV所以只有" } ], "id": 303, "review_annotation": "", "row_id": 303, "bounded_row_id": 303, "parent_row_ids": [ 303 ], "parent_row_id": 303, "packet_id": 22, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 606.1356826923078, "packet_end": 635.3236682692308, "raw_text": "再交给第35层投影", "selected_text": "再交给第35层投影", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 303 ], "packet_id": 22, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 621.0156826923078, "end": 622.5556826923078, "text": "借此获得KV信息", "raw_asr_text": "借此获得KV信息", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 22, "words": [ { "word": "借", "start": 621.0156826923078, "end": 621.3156826923077, "probability": 0.92919921875 }, { "word": "此", "start": 621.3156826923077, "end": 621.4956826923078, "probability": 1.0 }, { "word": "获", "start": 621.4956826923078, "end": 621.7356826923078, "probability": 1.0 }, { "word": "得", "start": 621.7356826923078, "end": 621.8756826923078, "probability": 0.998046875 }, { "word": "K", "start": 621.8756826923078, "end": 622.0556826923078, "probability": 0.9990234375 }, { "word": "V", "start": 622.0556826923078, "end": 622.1756826923078, "probability": 1.0 }, { "word": "信", "start": 622.1756826923078, "end": 622.3556826923078, "probability": 0.99853515625 }, { "word": "息", "start": 622.3556826923078, "end": 622.5556826923078, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09344960772801959, "compression_ratio": 1.206798866855524, "no_speech_prob": 7.747375657674027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投影成它的KV这里DeepSeek基本上是在质疑仅仅为了创建全局记忆库是否真的需要所有这些非线性计算于是直接砍掉中间的每一个串行流程使用该隐藏状态再交给第35层投影借此获得KV信息不过如果你还记得前面讲过的内容DeepSeek其实并不是每次都为每一层生成完全独立的庞大KV缓存CSAR已经通过三种不同模式让相邻层共享全局KV所以只有" } ], "id": 304, "review_annotation": "", "row_id": 304, "bounded_row_id": 304, "parent_row_ids": [ 304 ], "parent_row_id": 304, "packet_id": 22, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 606.1356826923078, "packet_end": 635.3236682692308, "raw_text": "借此获得KV信息", "selected_text": "借此获得KV信息", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 304 ], "packet_id": 22, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 622.5556826923078, "end": 625.1956826923077, "text": "不过如果你还记得前面讲过的内容", "raw_asr_text": "不过如果你还记得前面讲过的内容", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 22, "words": [ { "word": "不", "start": 622.5556826923078, "end": 622.9156826923078, "probability": 0.99560546875 }, { "word": "过", "start": 622.9156826923078, "end": 623.1356826923078, "probability": 1.0 }, { "word": "如果你", "start": 623.1356826923078, "end": 623.5156826923078, "probability": 0.98779296875 }, { "word": "还", "start": 623.5156826923078, "end": 623.7556826923078, "probability": 1.0 }, { "word": "记", "start": 623.7556826923078, "end": 623.9156826923078, "probability": 1.0 }, { "word": "得", "start": 623.9156826923078, "end": 624.0356826923078, "probability": 1.0 }, { "word": "前", "start": 624.0356826923078, "end": 624.1956826923077, "probability": 0.99951171875 }, { "word": "面", "start": 624.1956826923077, "end": 624.3556826923078, "probability": 0.99951171875 }, { "word": "讲", "start": 624.3556826923078, "end": 624.5556826923078, "probability": 0.99951171875 }, { "word": "过", "start": 624.5556826923078, "end": 624.7156826923078, "probability": 1.0 }, { "word": "的", "start": 624.7156826923078, "end": 624.8156826923077, "probability": 1.0 }, { "word": "内", "start": 624.8156826923077, "end": 624.9756826923078, "probability": 0.99951171875 }, { "word": "容", "start": 624.9756826923078, "end": 625.1956826923077, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09344960772801959, "compression_ratio": 1.206798866855524, "no_speech_prob": 7.747375657674027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投影成它的KV这里DeepSeek基本上是在质疑仅仅为了创建全局记忆库是否真的需要所有这些非线性计算于是直接砍掉中间的每一个串行流程使用该隐藏状态再交给第35层投影借此获得KV信息不过如果你还记得前面讲过的内容DeepSeek其实并不是每次都为每一层生成完全独立的庞大KV缓存CSAR已经通过三种不同模式让相邻层共享全局KV所以只有" } ], "id": 305, "review_annotation": "", "row_id": 305, "bounded_row_id": 305, "parent_row_ids": [ 305 ], "parent_row_id": 305, "packet_id": 22, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 606.1356826923078, "packet_end": 635.3236682692308, "raw_text": "不过如果你还记得前面讲过的内容", "selected_text": "不过如果你还记得前面讲过的内容", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 305 ], "packet_id": 22, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 625.1956826923077, "end": 627.7556826923078, "text": "DeepSeek其实并不是每次都为每一层生成", "raw_asr_text": "DeepSeek其实并不是每次都为每一层生成", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 22, "words": [ { "word": "De", "start": 625.1956826923077, "end": 625.4156826923078, "probability": 0.99072265625 }, { "word": "ep", "start": 625.4156826923078, "end": 625.5156826923078, "probability": 0.99853515625 }, { "word": "Se", "start": 625.5156826923078, "end": 625.6356826923078, "probability": 0.99267578125 }, { "word": "ek", "start": 625.6356826923078, "end": 625.7156826923078, "probability": 1.0 }, { "word": "其实", "start": 625.7156826923078, "end": 625.8956826923078, "probability": 0.96240234375 }, { "word": "并", "start": 625.8956826923078, "end": 626.1156826923078, "probability": 0.99951171875 }, { "word": "不是", "start": 626.1156826923078, "end": 626.2956826923078, "probability": 0.99951171875 }, { "word": "每", "start": 626.2956826923078, "end": 626.4556826923078, "probability": 1.0 }, { "word": "次", "start": 626.4556826923078, "end": 626.6356826923078, "probability": 0.99609375 }, { "word": "都", "start": 626.6356826923078, "end": 626.7756826923078, "probability": 0.9990234375 }, { "word": "为", "start": 626.7756826923078, "end": 626.9156826923078, "probability": 0.998046875 }, { "word": "每", "start": 626.9156826923078, "end": 627.0956826923078, "probability": 0.99951171875 }, { "word": "一", "start": 627.0956826923078, "end": 627.2556826923078, "probability": 0.9951171875 }, { "word": "层", "start": 627.2556826923078, "end": 627.3956826923078, "probability": 1.0 }, { "word": "生", "start": 627.3956826923078, "end": 627.5756826923079, "probability": 0.998046875 }, { "word": "成", "start": 627.5756826923079, "end": 627.7556826923078, "probability": 0.97607421875 } ], "asr_metrics": { "avg_logprob": -0.09344960772801959, "compression_ratio": 1.206798866855524, "no_speech_prob": 7.747375657674027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投影成它的KV这里DeepSeek基本上是在质疑仅仅为了创建全局记忆库是否真的需要所有这些非线性计算于是直接砍掉中间的每一个串行流程使用该隐藏状态再交给第35层投影借此获得KV信息不过如果你还记得前面讲过的内容DeepSeek其实并不是每次都为每一层生成完全独立的庞大KV缓存CSAR已经通过三种不同模式让相邻层共享全局KV所以只有" } ], "id": 306, "review_annotation": "", "row_id": 306, "bounded_row_id": 306, "parent_row_ids": [ 306 ], "parent_row_id": 306, "packet_id": 22, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 606.1356826923078, "packet_end": 635.3236682692308, "raw_text": "DeepSeek其实并不是每次都为每一层生成", "selected_text": "DeepSeek其实并不是每次都为每一层生成", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 306 ], "packet_id": 22, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 627.7556826923078, "end": 629.8556826923078, "text": "完全独立的庞大KV缓存", "raw_asr_text": "完全独立的庞大KV缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 22, "words": [ { "word": "完全", "start": 627.7556826923078, "end": 628.0556826923078, "probability": 0.99755859375 }, { "word": "独", "start": 628.0556826923078, "end": 628.2556826923078, "probability": 0.99951171875 }, { "word": "立", "start": 628.2556826923078, "end": 628.3956826923078, "probability": 1.0 }, { "word": "的", "start": 628.3956826923078, "end": 628.5156826923078, "probability": 0.99951171875 }, { "word": "庞", "start": 628.5156826923078, "end": 628.7556826923078, "probability": 0.93994140625 }, { "word": "大", "start": 628.7556826923078, "end": 628.9156826923078, "probability": 0.9951171875 }, { "word": "K", "start": 628.9156826923078, "end": 629.1356826923078, "probability": 0.99755859375 }, { "word": "V", "start": 629.1356826923078, "end": 629.3356826923078, "probability": 0.99853515625 }, { "word": "缓", "start": 629.3356826923078, "end": 629.5756826923079, "probability": 0.99658203125 }, { "word": "存", "start": 629.5756826923079, "end": 629.8556826923078, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09344960772801959, "compression_ratio": 1.206798866855524, "no_speech_prob": 7.747375657674027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投影成它的KV这里DeepSeek基本上是在质疑仅仅为了创建全局记忆库是否真的需要所有这些非线性计算于是直接砍掉中间的每一个串行流程使用该隐藏状态再交给第35层投影借此获得KV信息不过如果你还记得前面讲过的内容DeepSeek其实并不是每次都为每一层生成完全独立的庞大KV缓存CSAR已经通过三种不同模式让相邻层共享全局KV所以只有" } ], "id": 307, "review_annotation": "", "row_id": 307, "bounded_row_id": 307, "parent_row_ids": [ 307 ], "parent_row_id": 307, "packet_id": 22, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 606.1356826923078, "packet_end": 635.3236682692308, "raw_text": "完全独立的庞大KV缓存", "selected_text": "完全独立的庞大KV缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 307 ], "packet_id": 22, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 629.8556826923078, "end": 632.7556826923078, "text": "CSAR已经通过三种不同模式", "raw_asr_text": "CSAR已经通过三种不同模式", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 22, "words": [ { "word": "CS", "start": 629.8556826923078, "end": 630.5556826923078, "probability": 0.4892578125 }, { "word": "AR", "start": 630.5556826923078, "end": 630.8756826923078, "probability": 0.92724609375 }, { "word": "已经", "start": 630.8756826923078, "end": 631.2356826923078, "probability": 0.99609375 }, { "word": "通", "start": 631.2356826923078, "end": 631.4556826923078, "probability": 0.9990234375 }, { "word": "过", "start": 631.4556826923078, "end": 631.6156826923078, "probability": 1.0 }, { "word": "三", "start": 631.6156826923078, "end": 631.8156826923077, "probability": 0.984375 }, { "word": "种", "start": 631.8156826923077, "end": 632.0556826923078, "probability": 0.99951171875 }, { "word": "不同", "start": 632.0556826923078, "end": 632.2956826923078, "probability": 0.9990234375 }, { "word": "模", "start": 632.2956826923078, "end": 632.5356826923078, "probability": 0.998046875 }, { "word": "式", "start": 632.5356826923078, "end": 632.7556826923078, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09344960772801959, "compression_ratio": 1.206798866855524, "no_speech_prob": 7.747375657674027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投影成它的KV这里DeepSeek基本上是在质疑仅仅为了创建全局记忆库是否真的需要所有这些非线性计算于是直接砍掉中间的每一个串行流程使用该隐藏状态再交给第35层投影借此获得KV信息不过如果你还记得前面讲过的内容DeepSeek其实并不是每次都为每一层生成完全独立的庞大KV缓存CSAR已经通过三种不同模式让相邻层共享全局KV所以只有" } ], "id": 308, "review_annotation": "", "row_id": 308, "bounded_row_id": 308, "parent_row_ids": [ 308 ], "parent_row_id": 308, "packet_id": 22, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 606.1356826923078, "packet_end": 635.3236682692308, "raw_text": "CSAR已经通过三种不同模式", "selected_text": "CSAR已经通过三种不同模式", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 308 ], "packet_id": 22, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 632.7556826923078, "end": 634.5236682692308, "text": "让相邻层共享全局KV", "raw_asr_text": "让相邻层共享全局KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 22, "words": [ { "word": "让", "start": 632.7556826923078, "end": 632.9956826923078, "probability": 0.994140625 }, { "word": "相", "start": 632.9956826923078, "end": 633.1756826923078, "probability": 0.9560546875 }, { "word": "邻", "start": 633.1756826923078, "end": 633.3156826923077, "probability": 0.86328125 }, { "word": "层", "start": 633.3156826923077, "end": 633.5356826923078, "probability": 0.9990234375 }, { "word": "共", "start": 633.5356826923078, "end": 633.7756826923078, "probability": 0.998046875 }, { "word": "享", "start": 633.7756826923078, "end": 633.9956826923078, "probability": 0.998046875 }, { "word": "全", "start": 633.9956826923078, "end": 634.2556826923078, "probability": 0.99853515625 }, { "word": "局", "start": 634.2556826923078, "end": 634.4156826923078, "probability": 1.0 }, { "word": "K", "start": 634.4156826923078, "end": 634.6156826923078, "probability": 0.998046875 }, { "word": "V", "start": 634.6156826923078, "end": 634.7356826923078, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09344960772801959, "compression_ratio": 1.206798866855524, "no_speech_prob": 7.747375657674027e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "投影成它的KV这里DeepSeek基本上是在质疑仅仅为了创建全局记忆库是否真的需要所有这些非线性计算于是直接砍掉中间的每一个串行流程使用该隐藏状态再交给第35层投影借此获得KV信息不过如果你还记得前面讲过的内容DeepSeek其实并不是每次都为每一层生成完全独立的庞大KV缓存CSAR已经通过三种不同模式让相邻层共享全局KV所以只有" } ], "id": 309, "review_annotation": "", "row_id": 309, "bounded_row_id": 309, "parent_row_ids": [ 309 ], "parent_row_id": 309, "packet_id": 22, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 606.1356826923078, "packet_end": 635.3236682692308, "raw_text": "让相邻层共享全局KV", "selected_text": "让相邻层共享全局KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 309 ], "packet_id": 22, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 634.7436682692309, "end": 638.6236682692308, "text": "所以只有真正需要创建新全局KV的解码器层", "raw_asr_text": "所以只有真正需要创建新全局KV的解码器层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 23, "words": [ { "word": "所以", "start": 634.7436682692309, "end": 634.9436682692309, "probability": 0.88671875 }, { "word": "只有", "start": 634.9436682692309, "end": 635.3236682692309, "probability": 0.986328125 }, { "word": "真", "start": 635.3236682692309, "end": 635.5236682692308, "probability": 0.998046875 }, { "word": "正", "start": 635.5236682692308, "end": 635.7036682692309, "probability": 1.0 }, { "word": "需要", "start": 635.7036682692309, "end": 636.0236682692308, "probability": 0.99755859375 }, { "word": "创", "start": 636.0236682692308, "end": 636.3436682692309, "probability": 0.9931640625 }, { "word": "建", "start": 636.3436682692309, "end": 636.5436682692309, "probability": 0.99951171875 }, { "word": "新", "start": 636.5436682692309, "end": 636.7836682692308, "probability": 0.99609375 }, { "word": "全", "start": 636.7836682692308, "end": 637.0036682692308, "probability": 0.9990234375 }, { "word": "局", "start": 637.0036682692308, "end": 637.2036682692309, "probability": 1.0 }, { "word": "K", "start": 637.2036682692309, "end": 637.5236682692308, "probability": 0.98046875 }, { "word": "V", "start": 637.5236682692308, "end": 637.7436682692309, "probability": 0.99951171875 }, { "word": "的", "start": 637.7436682692309, "end": 637.9436682692309, "probability": 0.96923828125 }, { "word": "解", "start": 637.9436682692309, "end": 638.1036682692309, "probability": 0.994140625 }, { "word": "码", "start": 638.1036682692309, "end": 638.2636682692308, "probability": 0.9990234375 }, { "word": "器", "start": 638.2636682692308, "end": 638.3836682692308, "probability": 0.998046875 }, { "word": "层", "start": 638.3836682692308, "end": 638.6236682692308, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.06865988254547119, "compression_ratio": 1.29874213836478, "no_speech_prob": 7.476264052286297e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "共享全局KV所以只有真正需要创建新全局KV的解码器层才会用自身层专属的投影直接从最终编码器状态将其派生出来至于每一层具体使用哪种模式Deep Seek几乎是把这些模式硬编码进了层布局中在编码器中前两个仅使用SWA的层之后剩余18层按每6层一组划分一个完整层后面接5个复用层而在解码器中20层按每4层一组划分低足以一个玩意" } ], "id": 310, "review_annotation": "", "row_id": 310, "bounded_row_id": 310, "parent_row_ids": [ 310 ], "parent_row_id": 310, "packet_id": 23, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 633.7236682692309, "packet_end": 662.9116538461539, "raw_text": "所以只有真正需要创建新全局KV的解码器层", "selected_text": "所以只有真正需要创建新全局KV的解码器层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 310 ], "packet_id": 23, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 638.6236682692308, "end": 640.7036682692309, "text": "才会用自身层专属的投影", "raw_asr_text": "才会用自身层专属的投影", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 23, "words": [ { "word": "才", "start": 638.6236682692308, "end": 638.8836682692308, "probability": 0.9853515625 }, { "word": "会", "start": 638.8836682692308, "end": 639.0036682692308, "probability": 0.99755859375 }, { "word": "用", "start": 639.0036682692308, "end": 639.1436682692308, "probability": 0.9970703125 }, { "word": "自", "start": 639.1436682692308, "end": 639.3036682692309, "probability": 0.99951171875 }, { "word": "身", "start": 639.3036682692309, "end": 639.5036682692308, "probability": 0.9755859375 }, { "word": "层", "start": 639.5036682692308, "end": 639.7236682692309, "probability": 0.9970703125 }, { "word": "专", "start": 639.7236682692309, "end": 640.0636682692309, "probability": 0.9873046875 }, { "word": "属", "start": 640.0636682692309, "end": 640.2636682692308, "probability": 0.99365234375 }, { "word": "的", "start": 640.2636682692308, "end": 640.3636682692309, "probability": 0.9833984375 }, { "word": "投", "start": 640.3636682692309, "end": 640.5236682692308, "probability": 0.99951171875 }, { "word": "影", "start": 640.5236682692308, "end": 640.7036682692309, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.06865988254547119, "compression_ratio": 1.29874213836478, "no_speech_prob": 7.476264052286297e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "共享全局KV所以只有真正需要创建新全局KV的解码器层才会用自身层专属的投影直接从最终编码器状态将其派生出来至于每一层具体使用哪种模式Deep Seek几乎是把这些模式硬编码进了层布局中在编码器中前两个仅使用SWA的层之后剩余18层按每6层一组划分一个完整层后面接5个复用层而在解码器中20层按每4层一组划分低足以一个玩意" } ], "id": 311, "review_annotation": "", "row_id": 311, "bounded_row_id": 311, "parent_row_ids": [ 311 ], "parent_row_id": 311, "packet_id": 23, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 633.7236682692309, "packet_end": 662.9116538461539, "raw_text": "才会用自身层专属的投影", "selected_text": "才会用自身层专属的投影", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 311 ], "packet_id": 23, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 640.7036682692309, "end": 643.6836682692309, "text": "直接从最终编码器状态将其派生出来", "raw_asr_text": "直接从最终编码器状态将其派生出来", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 23, "words": [ { "word": "直接", "start": 640.7036682692309, "end": 641.0636682692309, "probability": 0.98583984375 }, { "word": "从", "start": 641.0636682692309, "end": 641.3436682692309, "probability": 1.0 }, { "word": "最", "start": 641.3436682692309, "end": 641.5836682692309, "probability": 1.0 }, { "word": "终", "start": 641.5836682692309, "end": 641.7236682692309, "probability": 0.998046875 }, { "word": "编", "start": 641.7236682692309, "end": 641.8636682692309, "probability": 0.998046875 }, { "word": "码", "start": 641.8636682692309, "end": 642.0036682692308, "probability": 1.0 }, { "word": "器", "start": 642.0036682692308, "end": 642.1236682692308, "probability": 0.99951171875 }, { "word": "状", "start": 642.1236682692308, "end": 642.2636682692308, "probability": 0.99755859375 }, { "word": "态", "start": 642.2636682692308, "end": 642.4836682692309, "probability": 0.9990234375 }, { "word": "将", "start": 642.4836682692309, "end": 642.7036682692309, "probability": 0.99658203125 }, { "word": "其", "start": 642.7036682692309, "end": 643.0036682692308, "probability": 0.99951171875 }, { "word": "派", "start": 643.0036682692308, "end": 643.2236682692309, "probability": 0.9970703125 }, { "word": "生", "start": 643.2236682692309, "end": 643.3836682692308, "probability": 0.89111328125 }, { "word": "出来", "start": 643.3836682692308, "end": 643.6836682692309, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.06865988254547119, "compression_ratio": 1.29874213836478, "no_speech_prob": 7.476264052286297e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "共享全局KV所以只有真正需要创建新全局KV的解码器层才会用自身层专属的投影直接从最终编码器状态将其派生出来至于每一层具体使用哪种模式Deep Seek几乎是把这些模式硬编码进了层布局中在编码器中前两个仅使用SWA的层之后剩余18层按每6层一组划分一个完整层后面接5个复用层而在解码器中20层按每4层一组划分低足以一个玩意" } ], "id": 312, "review_annotation": "", "row_id": 312, "bounded_row_id": 312, "parent_row_ids": [ 312 ], "parent_row_id": 312, "packet_id": 23, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 633.7236682692309, "packet_end": 662.9116538461539, "raw_text": "直接从最终编码器状态将其派生出来", "selected_text": "直接从最终编码器状态将其派生出来", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 312 ], "packet_id": 23, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 643.6836682692309, "end": 646.1636682692309, "text": "至于每一层具体使用哪种模式", "raw_asr_text": "至于每一层具体使用哪种模式", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 23, "words": [ { "word": "至", "start": 643.6836682692309, "end": 644.0036682692308, "probability": 0.99951171875 }, { "word": "于", "start": 644.0036682692308, "end": 644.1836682692309, "probability": 1.0 }, { "word": "每", "start": 644.1836682692309, "end": 644.4836682692309, "probability": 0.9990234375 }, { "word": "一", "start": 644.4836682692309, "end": 644.6636682692309, "probability": 0.99853515625 }, { "word": "层", "start": 644.6636682692309, "end": 644.8236682692309, "probability": 1.0 }, { "word": "具", "start": 644.8236682692309, "end": 645.0036682692308, "probability": 0.99951171875 }, { "word": "体", "start": 645.0036682692308, "end": 645.1636682692309, "probability": 1.0 }, { "word": "使", "start": 645.1636682692309, "end": 645.3636682692309, "probability": 0.9951171875 }, { "word": "用", "start": 645.3636682692309, "end": 645.4836682692309, "probability": 1.0 }, { "word": "哪", "start": 645.4836682692309, "end": 645.6036682692309, "probability": 0.9794921875 }, { "word": "种", "start": 645.6036682692309, "end": 645.7436682692309, "probability": 0.9951171875 }, { "word": "模", "start": 645.7436682692309, "end": 645.9236682692309, "probability": 0.99951171875 }, { "word": "式", "start": 645.9236682692309, "end": 646.1636682692309, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.06865988254547119, "compression_ratio": 1.29874213836478, "no_speech_prob": 7.476264052286297e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "共享全局KV所以只有真正需要创建新全局KV的解码器层才会用自身层专属的投影直接从最终编码器状态将其派生出来至于每一层具体使用哪种模式Deep Seek几乎是把这些模式硬编码进了层布局中在编码器中前两个仅使用SWA的层之后剩余18层按每6层一组划分一个完整层后面接5个复用层而在解码器中20层按每4层一组划分低足以一个玩意" } ], "id": 313, "review_annotation": "", "row_id": 313, "bounded_row_id": 313, "parent_row_ids": [ 313 ], "parent_row_id": 313, "packet_id": 23, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 633.7236682692309, "packet_end": 662.9116538461539, "raw_text": "至于每一层具体使用哪种模式", "selected_text": "至于每一层具体使用哪种模式", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 313 ], "packet_id": 23, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 646.1636682692309, "end": 649.7236682692309, "text": "Deep Seek几乎是把这些模式硬编码进了层布局中", "raw_asr_text": "Deep Seek几乎是把这些模式硬编码进了层布局中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 23, "words": [ { "word": "De", "start": 646.1636682692309, "end": 646.4036682692308, "probability": 0.904296875 }, { "word": "ep", "start": 646.4036682692308, "end": 646.5036682692308, "probability": 0.97607421875 }, { "word": " Se", "start": 646.5036682692308, "end": 646.6636682692309, "probability": 0.33837890625 }, { "word": "ek", "start": 646.6636682692309, "end": 646.8036682692309, "probability": 0.98583984375 }, { "word": "几", "start": 646.8036682692309, "end": 647.1636682692309, "probability": 0.998046875 }, { "word": "乎", "start": 647.1636682692309, "end": 647.2636682692308, "probability": 1.0 }, { "word": "是", "start": 647.2636682692308, "end": 647.4036682692308, "probability": 0.998046875 }, { "word": "把", "start": 647.4036682692308, "end": 647.5236682692308, "probability": 0.99951171875 }, { "word": "这", "start": 647.5236682692308, "end": 647.6636682692309, "probability": 0.99951171875 }, { "word": "些", "start": 647.6636682692309, "end": 647.7836682692308, "probability": 0.99951171875 }, { "word": "模", "start": 647.7836682692308, "end": 647.9636682692309, "probability": 0.99951171875 }, { "word": "式", "start": 647.9636682692309, "end": 648.1436682692308, "probability": 1.0 }, { "word": "硬", "start": 648.1436682692308, "end": 648.3636682692309, "probability": 0.974609375 }, { "word": "编", "start": 648.3636682692309, "end": 648.5436682692309, "probability": 0.9990234375 }, { "word": "码", "start": 648.5436682692309, "end": 648.7436682692309, "probability": 1.0 }, { "word": "进", "start": 648.7436682692309, "end": 648.9036682692308, "probability": 0.99755859375 }, { "word": "了", "start": 648.9036682692308, "end": 649.0436682692309, "probability": 0.99853515625 }, { "word": "层", "start": 649.0436682692309, "end": 649.2036682692309, "probability": 0.99951171875 }, { "word": "布", "start": 649.2036682692309, "end": 649.3436682692309, "probability": 0.9873046875 }, { "word": "局", "start": 649.3436682692309, "end": 649.5236682692308, "probability": 0.99951171875 }, { "word": "中", "start": 649.5236682692308, "end": 649.7236682692309, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.06865988254547119, "compression_ratio": 1.29874213836478, "no_speech_prob": 7.476264052286297e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "共享全局KV所以只有真正需要创建新全局KV的解码器层才会用自身层专属的投影直接从最终编码器状态将其派生出来至于每一层具体使用哪种模式Deep Seek几乎是把这些模式硬编码进了层布局中在编码器中前两个仅使用SWA的层之后剩余18层按每6层一组划分一个完整层后面接5个复用层而在解码器中20层按每4层一组划分低足以一个玩意" } ], "id": 314, "review_annotation": "", "row_id": 314, "bounded_row_id": 314, "parent_row_ids": [ 314 ], "parent_row_id": 314, "packet_id": 23, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 633.7236682692309, "packet_end": 662.9116538461539, "raw_text": "Deep Seek几乎是把这些模式硬编码进了层布局中", "selected_text": "Deep Seek几乎是把这些模式硬编码进了层布局中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 314 ], "packet_id": 23, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 649.7236682692309, "end": 650.7836682692308, "text": "在编码器中", "raw_asr_text": "在编码器中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 23, "words": [ { "word": "在", "start": 649.7236682692309, "end": 650.0836682692309, "probability": 0.99853515625 }, { "word": "编", "start": 650.0836682692309, "end": 650.3036682692309, "probability": 0.99951171875 }, { "word": "码", "start": 650.3036682692309, "end": 650.4436682692309, "probability": 1.0 }, { "word": "器", "start": 650.4436682692309, "end": 650.5836682692309, "probability": 1.0 }, { "word": "中", "start": 650.5836682692309, "end": 650.7836682692308, "probability": 0.99462890625 } ], "asr_metrics": { "avg_logprob": -0.06865988254547119, "compression_ratio": 1.29874213836478, "no_speech_prob": 7.476264052286297e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "共享全局KV所以只有真正需要创建新全局KV的解码器层才会用自身层专属的投影直接从最终编码器状态将其派生出来至于每一层具体使用哪种模式Deep Seek几乎是把这些模式硬编码进了层布局中在编码器中前两个仅使用SWA的层之后剩余18层按每6层一组划分一个完整层后面接5个复用层而在解码器中20层按每4层一组划分低足以一个玩意" } ], "id": 315, "review_annotation": "", "row_id": 315, "bounded_row_id": 315, "parent_row_ids": [ 315 ], "parent_row_id": 315, "packet_id": 23, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 633.7236682692309, "packet_end": 662.9116538461539, "raw_text": "在编码器中", "selected_text": "在编码器中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 315 ], "packet_id": 23, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 650.7836682692308, "end": 653.3836682692308, "text": "前两个仅使用SWA的层之后", "raw_asr_text": "前两个仅使用SWA的层之后", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 23, "words": [ { "word": "前", "start": 650.7836682692308, "end": 651.1036682692309, "probability": 0.97705078125 }, { "word": "两", "start": 651.1036682692309, "end": 651.2436682692309, "probability": 0.99853515625 }, { "word": "个", "start": 651.2436682692309, "end": 651.3836682692308, "probability": 0.9990234375 }, { "word": "仅", "start": 651.3836682692308, "end": 651.5436682692309, "probability": 0.9990234375 }, { "word": "使", "start": 651.5436682692309, "end": 651.7236682692309, "probability": 0.99462890625 }, { "word": "用", "start": 651.7236682692309, "end": 651.9036682692308, "probability": 1.0 }, { "word": "S", "start": 651.9036682692308, "end": 652.1036682692309, "probability": 0.97509765625 }, { "word": "WA", "start": 652.1036682692309, "end": 652.5636682692309, "probability": 0.98046875 }, { "word": "的", "start": 652.5636682692309, "end": 652.7236682692309, "probability": 0.9921875 }, { "word": "层", "start": 652.7236682692309, "end": 652.9636682692309, "probability": 0.99951171875 }, { "word": "之", "start": 652.9636682692309, "end": 653.1636682692309, "probability": 0.998046875 }, { "word": "后", "start": 653.1636682692309, "end": 653.3836682692308, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.06865988254547119, "compression_ratio": 1.29874213836478, "no_speech_prob": 7.476264052286297e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "共享全局KV所以只有真正需要创建新全局KV的解码器层才会用自身层专属的投影直接从最终编码器状态将其派生出来至于每一层具体使用哪种模式Deep Seek几乎是把这些模式硬编码进了层布局中在编码器中前两个仅使用SWA的层之后剩余18层按每6层一组划分一个完整层后面接5个复用层而在解码器中20层按每4层一组划分低足以一个玩意" } ], "id": 316, "review_annotation": "", "row_id": 316, "bounded_row_id": 316, "parent_row_ids": [ 316 ], "parent_row_id": 316, "packet_id": 23, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 633.7236682692309, "packet_end": 662.9116538461539, "raw_text": "前两个仅使用SWA的层之后", "selected_text": "前两个仅使用SWA的层之后", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 316 ], "packet_id": 23, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 653.3836682692308, "end": 656.0036682692308, "text": "剩余18层按每6层一组划分", "raw_asr_text": "剩余18层按每6层一组划分", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 23, "words": [ { "word": "剩", "start": 653.3836682692308, "end": 653.6836682692309, "probability": 0.99853515625 }, { "word": "余", "start": 653.6836682692309, "end": 653.8436682692309, "probability": 0.9931640625 }, { "word": "18", "start": 653.8436682692309, "end": 654.1036682692309, "probability": 0.94287109375 }, { "word": "层", "start": 654.1036682692309, "end": 654.4036682692308, "probability": 0.99951171875 }, { "word": "按", "start": 654.4036682692308, "end": 654.6636682692309, "probability": 0.99462890625 }, { "word": "每", "start": 654.6636682692309, "end": 654.9436682692309, "probability": 0.99755859375 }, { "word": "6", "start": 654.9436682692309, "end": 655.1236682692308, "probability": 0.9169921875 }, { "word": "层", "start": 655.1236682692308, "end": 655.3236682692309, "probability": 1.0 }, { "word": "一", "start": 655.3236682692309, "end": 655.5036682692308, "probability": 0.99560546875 }, { "word": "组", "start": 655.5036682692308, "end": 655.6436682692308, "probability": 0.99609375 }, { "word": "划", "start": 655.6436682692308, "end": 655.8036682692309, "probability": 0.9951171875 }, { "word": "分", "start": 655.8036682692309, "end": 656.0036682692308, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.06865988254547119, "compression_ratio": 1.29874213836478, "no_speech_prob": 7.476264052286297e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "共享全局KV所以只有真正需要创建新全局KV的解码器层才会用自身层专属的投影直接从最终编码器状态将其派生出来至于每一层具体使用哪种模式Deep Seek几乎是把这些模式硬编码进了层布局中在编码器中前两个仅使用SWA的层之后剩余18层按每6层一组划分一个完整层后面接5个复用层而在解码器中20层按每4层一组划分低足以一个玩意" } ], "id": 317, "review_annotation": "", "row_id": 317, "bounded_row_id": 317, "parent_row_ids": [ 317 ], "parent_row_id": 317, "packet_id": 23, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 633.7236682692309, "packet_end": 662.9116538461539, "raw_text": "剩余18层按每6层一组划分", "selected_text": "剩余18层按每6层一组划分", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 317 ], "packet_id": 23, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 656.0036682692308, "end": 658.2836682692308, "text": "一个完整层后面接5个复用层", "raw_asr_text": "一个完整层后面接5个复用层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 23, "words": [ { "word": "一个", "start": 656.0036682692308, "end": 656.3236682692309, "probability": 0.98046875 }, { "word": "完", "start": 656.3236682692309, "end": 656.5236682692308, "probability": 0.99951171875 }, { "word": "整", "start": 656.5236682692308, "end": 656.6436682692308, "probability": 1.0 }, { "word": "层", "start": 656.6436682692308, "end": 656.8636682692309, "probability": 1.0 }, { "word": "后", "start": 656.8636682692309, "end": 657.1036682692309, "probability": 0.890625 }, { "word": "面", "start": 657.1036682692309, "end": 657.2836682692308, "probability": 0.99951171875 }, { "word": "接", "start": 657.2836682692308, "end": 657.4636682692309, "probability": 0.931640625 }, { "word": "5", "start": 657.4636682692309, "end": 657.6636682692309, "probability": 0.95068359375 }, { "word": "个", "start": 657.6636682692309, "end": 657.8236682692309, "probability": 0.99755859375 }, { "word": "复", "start": 657.8236682692309, "end": 657.9436682692309, "probability": 0.884765625 }, { "word": "用", "start": 657.9436682692309, "end": 658.0436682692309, "probability": 0.9990234375 }, { "word": "层", "start": 658.0436682692309, "end": 658.2836682692308, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.06865988254547119, "compression_ratio": 1.29874213836478, "no_speech_prob": 7.476264052286297e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "共享全局KV所以只有真正需要创建新全局KV的解码器层才会用自身层专属的投影直接从最终编码器状态将其派生出来至于每一层具体使用哪种模式Deep Seek几乎是把这些模式硬编码进了层布局中在编码器中前两个仅使用SWA的层之后剩余18层按每6层一组划分一个完整层后面接5个复用层而在解码器中20层按每4层一组划分低足以一个玩意" } ], "id": 318, "review_annotation": "", "row_id": 318, "bounded_row_id": 318, "parent_row_ids": [ 318 ], "parent_row_id": 318, "packet_id": 23, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 633.7236682692309, "packet_end": 662.9116538461539, "raw_text": "一个完整层后面接5个复用层", "selected_text": "一个完整层后面接5个复用层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 318 ], "packet_id": 23, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 658.2836682692308, "end": 661.6436682692308, "text": "而在解码器中20层按每4层一组划分", "raw_asr_text": "而在解码器中20层按每4层一组划分", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 23, "words": [ { "word": "而", "start": 658.2836682692308, "end": 658.6036682692309, "probability": 0.99755859375 }, { "word": "在", "start": 658.6036682692309, "end": 658.8036682692309, "probability": 0.99853515625 }, { "word": "解", "start": 658.8036682692309, "end": 658.9636682692309, "probability": 0.99853515625 }, { "word": "码", "start": 658.9636682692309, "end": 659.1436682692308, "probability": 1.0 }, { "word": "器", "start": 659.1436682692308, "end": 659.2836682692308, "probability": 0.99951171875 }, { "word": "中", "start": 659.2836682692308, "end": 659.4436682692309, "probability": 0.9990234375 }, { "word": "20", "start": 659.4436682692309, "end": 659.8236682692309, "probability": 0.99462890625 }, { "word": "层", "start": 659.8236682692309, "end": 660.0836682692309, "probability": 0.99951171875 }, { "word": "按", "start": 660.0836682692309, "end": 660.3036682692309, "probability": 0.99609375 }, { "word": "每", "start": 660.3036682692309, "end": 660.4836682692309, "probability": 0.99853515625 }, { "word": "4", "start": 660.4836682692309, "end": 660.6636682692309, "probability": 0.99267578125 }, { "word": "层", "start": 660.6636682692309, "end": 660.8436682692309, "probability": 0.9990234375 }, { "word": "一", "start": 660.8436682692309, "end": 661.1236682692308, "probability": 0.998046875 }, { "word": "组", "start": 661.1236682692308, "end": 661.2636682692308, "probability": 1.0 }, { "word": "划", "start": 661.2636682692308, "end": 661.4436682692309, "probability": 0.9990234375 }, { "word": "分", "start": 661.4436682692309, "end": 661.6436682692308, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.06865988254547119, "compression_ratio": 1.29874213836478, "no_speech_prob": 7.476264052286297e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "共享全局KV所以只有真正需要创建新全局KV的解码器层才会用自身层专属的投影直接从最终编码器状态将其派生出来至于每一层具体使用哪种模式Deep Seek几乎是把这些模式硬编码进了层布局中在编码器中前两个仅使用SWA的层之后剩余18层按每6层一组划分一个完整层后面接5个复用层而在解码器中20层按每4层一组划分低足以一个玩意" } ], "id": 319, "review_annotation": "", "row_id": 319, "bounded_row_id": 319, "parent_row_ids": [ 319 ], "parent_row_id": 319, "packet_id": 23, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 633.7236682692309, "packet_end": 662.9116538461539, "raw_text": "而在解码器中20层按每4层一组划分", "selected_text": "而在解码器中20层按每4层一组划分", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 319 ], "packet_id": 23, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 662.1116538461539, "end": 663.7116538461539, "text": "D组以一个完整层开头", "raw_asr_text": "D组以一个完整层开头", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 24, "words": [ { "word": "D", "start": 661.311653846154, "end": 662.031653846154, "probability": 0.10186767578125 }, { "word": "组", "start": 662.031653846154, "end": 662.291653846154, "probability": 0.97607421875 }, { "word": "以", "start": 662.291653846154, "end": 662.511653846154, "probability": 0.95361328125 }, { "word": "一个", "start": 662.511653846154, "end": 662.7316538461539, "probability": 0.9765625 }, { "word": "完", "start": 662.7316538461539, "end": 662.9516538461539, "probability": 0.990234375 }, { "word": "整", "start": 662.9516538461539, "end": 663.151653846154, "probability": 1.0 }, { "word": "层", "start": 663.151653846154, "end": 663.311653846154, "probability": 0.9990234375 }, { "word": "开", "start": 663.311653846154, "end": 663.511653846154, "probability": 0.99560546875 }, { "word": "头", "start": 663.511653846154, "end": 663.7116538461539, "probability": 0.99658203125 } ], "asr_metrics": { "avg_logprob": -0.1042110481072421, "compression_ratio": 1.3841059602649006, "no_speech_prob": 3.177028431999851e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "D组以一个完整层开头后面接三个复用层后面的各组则以一个众索引层开头再接三个复用层所以综合起来看你基本上同时获得了两个层面的成本节省首先CD避免了仅仅为了生成解码器全局KV就让整个预填充跑变模型上半部分相反它直接从最终编码器表示派生这份KV然后借助CSAR一旦某个解码器层生成了这份KV接下来的几层" } ], "id": 320, "review_annotation": "", "row_id": 320, "bounded_row_id": 320, "parent_row_ids": [ 320 ], "parent_row_id": 320, "packet_id": 24, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 661.311653846154, "packet_end": 690.499639423077, "raw_text": "D组以一个完整层开头", "selected_text": "D组以一个完整层开头", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 320 ], "packet_id": 24, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 663.7116538461539, "end": 665.2316538461539, "text": "后面接三个复用层", "raw_asr_text": "后面接三个复用层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 24, "words": [ { "word": "后", "start": 663.7116538461539, "end": 663.9716538461539, "probability": 0.8798828125 }, { "word": "面", "start": 663.9716538461539, "end": 664.171653846154, "probability": 0.99853515625 }, { "word": "接", "start": 664.171653846154, "end": 664.311653846154, "probability": 0.9580078125 }, { "word": "三", "start": 664.311653846154, "end": 664.511653846154, "probability": 0.81298828125 }, { "word": "个", "start": 664.511653846154, "end": 664.6916538461539, "probability": 0.99755859375 }, { "word": "复", "start": 664.6916538461539, "end": 664.8316538461539, "probability": 0.853515625 }, { "word": "用", "start": 664.8316538461539, "end": 664.9716538461539, "probability": 0.99853515625 }, { "word": "层", "start": 664.9716538461539, "end": 665.2316538461539, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.1042110481072421, "compression_ratio": 1.3841059602649006, "no_speech_prob": 3.177028431999851e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "D组以一个完整层开头后面接三个复用层后面的各组则以一个众索引层开头再接三个复用层所以综合起来看你基本上同时获得了两个层面的成本节省首先CD避免了仅仅为了生成解码器全局KV就让整个预填充跑变模型上半部分相反它直接从最终编码器表示派生这份KV然后借助CSAR一旦某个解码器层生成了这份KV接下来的几层" } ], "id": 321, "review_annotation": "", "row_id": 321, "bounded_row_id": 321, "parent_row_ids": [ 321 ], "parent_row_id": 321, "packet_id": 24, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 661.311653846154, "packet_end": 690.499639423077, "raw_text": "后面接三个复用层", "selected_text": "后面接三个复用层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 321 ], "packet_id": 24, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 665.2316538461539, "end": 667.911653846154, "text": "后面的各组则以一个众索引层开头", "raw_asr_text": "后面的各组则以一个众索引层开头", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 24, "words": [ { "word": "后", "start": 665.2316538461539, "end": 665.511653846154, "probability": 0.9501953125 }, { "word": "面", "start": 665.511653846154, "end": 665.671653846154, "probability": 0.99951171875 }, { "word": "的", "start": 665.671653846154, "end": 665.771653846154, "probability": 0.98486328125 }, { "word": "各", "start": 665.771653846154, "end": 665.931653846154, "probability": 0.958984375 }, { "word": "组", "start": 665.931653846154, "end": 666.171653846154, "probability": 0.9970703125 }, { "word": "则", "start": 666.171653846154, "end": 666.391653846154, "probability": 0.9912109375 }, { "word": "以", "start": 666.391653846154, "end": 666.5716538461539, "probability": 0.99072265625 }, { "word": "一个", "start": 666.5716538461539, "end": 666.811653846154, "probability": 0.99072265625 }, { "word": "众", "start": 666.811653846154, "end": 667.051653846154, "probability": 0.8203125 }, { "word": "索", "start": 667.051653846154, "end": 667.2116538461539, "probability": 0.96484375 }, { "word": "引", "start": 667.2116538461539, "end": 667.3516538461539, "probability": 0.66650390625 }, { "word": "层", "start": 667.3516538461539, "end": 667.511653846154, "probability": 0.939453125 }, { "word": "开", "start": 667.511653846154, "end": 667.6916538461539, "probability": 0.99560546875 }, { "word": "头", "start": 667.6916538461539, "end": 667.911653846154, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.1042110481072421, "compression_ratio": 1.3841059602649006, "no_speech_prob": 3.177028431999851e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "D组以一个完整层开头后面接三个复用层后面的各组则以一个众索引层开头再接三个复用层所以综合起来看你基本上同时获得了两个层面的成本节省首先CD避免了仅仅为了生成解码器全局KV就让整个预填充跑变模型上半部分相反它直接从最终编码器表示派生这份KV然后借助CSAR一旦某个解码器层生成了这份KV接下来的几层" } ], "id": 322, "review_annotation": "", "row_id": 322, "bounded_row_id": 322, "parent_row_ids": [ 322 ], "parent_row_id": 322, "packet_id": 24, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 661.311653846154, "packet_end": 690.499639423077, "raw_text": "后面的各组则以一个众索引层开头", "selected_text": "后面的各组则以一个众索引层开头", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 322 ], "packet_id": 24, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 667.911653846154, "end": 669.4916538461539, "text": "再接三个复用层", "raw_asr_text": "再接三个复用层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 24, "words": [ { "word": "再", "start": 667.911653846154, "end": 668.271653846154, "probability": 0.9755859375 }, { "word": "接", "start": 668.271653846154, "end": 668.531653846154, "probability": 0.99658203125 }, { "word": "三", "start": 668.531653846154, "end": 668.7316538461539, "probability": 0.98681640625 }, { "word": "个", "start": 668.7316538461539, "end": 668.891653846154, "probability": 0.9990234375 }, { "word": "复", "start": 668.891653846154, "end": 669.051653846154, "probability": 0.9990234375 }, { "word": "用", "start": 669.051653846154, "end": 669.2316538461539, "probability": 1.0 }, { "word": "层", "start": 669.2316538461539, "end": 669.4916538461539, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.1042110481072421, "compression_ratio": 1.3841059602649006, "no_speech_prob": 3.177028431999851e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "D组以一个完整层开头后面接三个复用层后面的各组则以一个众索引层开头再接三个复用层所以综合起来看你基本上同时获得了两个层面的成本节省首先CD避免了仅仅为了生成解码器全局KV就让整个预填充跑变模型上半部分相反它直接从最终编码器表示派生这份KV然后借助CSAR一旦某个解码器层生成了这份KV接下来的几层" } ], "id": 323, "review_annotation": "", "row_id": 323, "bounded_row_id": 323, "parent_row_ids": [ 323 ], "parent_row_id": 323, "packet_id": 24, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 661.311653846154, "packet_end": 690.499639423077, "raw_text": "再接三个复用层", "selected_text": "再接三个复用层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 323 ], "packet_id": 24, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 669.4916538461539, "end": 670.671653846154, "text": "所以综合起来看", "raw_asr_text": "所以综合起来看", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 24, "words": [ { "word": "所以", "start": 669.4916538461539, "end": 669.811653846154, "probability": 0.98583984375 }, { "word": "综", "start": 669.811653846154, "end": 670.031653846154, "probability": 0.998046875 }, { "word": "合", "start": 670.031653846154, "end": 670.1916538461539, "probability": 0.99951171875 }, { "word": "起", "start": 670.1916538461539, "end": 670.3316538461539, "probability": 0.998046875 }, { "word": "来", "start": 670.3316538461539, "end": 670.4716538461539, "probability": 1.0 }, { "word": "看", "start": 670.4716538461539, "end": 670.671653846154, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1042110481072421, "compression_ratio": 1.3841059602649006, "no_speech_prob": 3.177028431999851e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "D组以一个完整层开头后面接三个复用层后面的各组则以一个众索引层开头再接三个复用层所以综合起来看你基本上同时获得了两个层面的成本节省首先CD避免了仅仅为了生成解码器全局KV就让整个预填充跑变模型上半部分相反它直接从最终编码器表示派生这份KV然后借助CSAR一旦某个解码器层生成了这份KV接下来的几层" } ], "id": 324, "review_annotation": "", "row_id": 324, "bounded_row_id": 324, "parent_row_ids": [ 324 ], "parent_row_id": 324, "packet_id": 24, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 661.311653846154, "packet_end": 690.499639423077, "raw_text": "所以综合起来看", "selected_text": "所以综合起来看", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 324 ], "packet_id": 24, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 670.671653846154, "end": 674.011653846154, "text": "你基本上同时获得了两个层面的成本节省", "raw_asr_text": "你基本上同时获得了两个层面的成本节省", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 24, "words": [ { "word": "你", "start": 670.671653846154, "end": 670.931653846154, "probability": 0.97705078125 }, { "word": "基本", "start": 670.931653846154, "end": 671.2116538461539, "probability": 0.9970703125 }, { "word": "上", "start": 671.2116538461539, "end": 671.411653846154, "probability": 0.99755859375 }, { "word": "同", "start": 671.411653846154, "end": 671.5916538461539, "probability": 0.9951171875 }, { "word": "时", "start": 671.5916538461539, "end": 671.771653846154, "probability": 0.9990234375 }, { "word": "获", "start": 671.771653846154, "end": 672.011653846154, "probability": 0.9990234375 }, { "word": "得", "start": 672.011653846154, "end": 672.151653846154, "probability": 0.99951171875 }, { "word": "了", "start": 672.151653846154, "end": 672.291653846154, "probability": 0.99267578125 }, { "word": "两", "start": 672.291653846154, "end": 672.531653846154, "probability": 0.9736328125 }, { "word": "个", "start": 672.531653846154, "end": 672.7316538461539, "probability": 0.9931640625 }, { "word": "层", "start": 672.7316538461539, "end": 672.8716538461539, "probability": 1.0 }, { "word": "面", "start": 672.8716538461539, "end": 673.051653846154, "probability": 0.99951171875 }, { "word": "的", "start": 673.051653846154, "end": 673.171653846154, "probability": 0.9970703125 }, { "word": "成", "start": 673.171653846154, "end": 673.3516538461539, "probability": 0.99951171875 }, { "word": "本", "start": 673.3516538461539, "end": 673.551653846154, "probability": 0.9990234375 }, { "word": "节", "start": 673.551653846154, "end": 673.7316538461539, "probability": 0.98876953125 }, { "word": "省", "start": 673.7316538461539, "end": 674.011653846154, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.1042110481072421, "compression_ratio": 1.3841059602649006, "no_speech_prob": 3.177028431999851e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "D组以一个完整层开头后面接三个复用层后面的各组则以一个众索引层开头再接三个复用层所以综合起来看你基本上同时获得了两个层面的成本节省首先CD避免了仅仅为了生成解码器全局KV就让整个预填充跑变模型上半部分相反它直接从最终编码器表示派生这份KV然后借助CSAR一旦某个解码器层生成了这份KV接下来的几层" } ], "id": 325, "review_annotation": "", "row_id": 325, "bounded_row_id": 325, "parent_row_ids": [ 325 ], "parent_row_id": 325, "packet_id": 24, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 661.311653846154, "packet_end": 690.499639423077, "raw_text": "你基本上同时获得了两个层面的成本节省", "selected_text": "你基本上同时获得了两个层面的成本节省", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 325 ], "packet_id": 24, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 674.011653846154, "end": 674.5716538461539, "text": "首先", "raw_asr_text": "首先", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 24, "words": [ { "word": "首先", "start": 674.011653846154, "end": 674.5716538461539, "probability": 0.99365234375 } ], "asr_metrics": { "avg_logprob": -0.1042110481072421, "compression_ratio": 1.3841059602649006, "no_speech_prob": 3.177028431999851e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "D组以一个完整层开头后面接三个复用层后面的各组则以一个众索引层开头再接三个复用层所以综合起来看你基本上同时获得了两个层面的成本节省首先CD避免了仅仅为了生成解码器全局KV就让整个预填充跑变模型上半部分相反它直接从最终编码器表示派生这份KV然后借助CSAR一旦某个解码器层生成了这份KV接下来的几层" } ], "id": 326, "review_annotation": "", "row_id": 326, "bounded_row_id": 326, "parent_row_ids": [ 326 ], "parent_row_id": 326, "packet_id": 24, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 661.311653846154, "packet_end": 690.499639423077, "raw_text": "首先", "selected_text": "首先", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 326 ], "packet_id": 24, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 674.5716538461539, "end": 678.3316538461539, "text": "CD避免了仅仅为了生成解码器全局KV", "raw_asr_text": "CD避免了仅仅为了生成解码器全局KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 24, "words": [ { "word": "CD", "start": 674.5716538461539, "end": 675.271653846154, "probability": 0.55224609375 }, { "word": "避", "start": 675.271653846154, "end": 675.6116538461539, "probability": 0.9931640625 }, { "word": "免", "start": 675.6116538461539, "end": 675.791653846154, "probability": 1.0 }, { "word": "了", "start": 675.791653846154, "end": 675.911653846154, "probability": 0.7724609375 }, { "word": "仅", "start": 675.911653846154, "end": 676.1116538461539, "probability": 0.99755859375 }, { "word": "仅", "start": 676.1116538461539, "end": 676.2116538461539, "probability": 0.9990234375 }, { "word": "为", "start": 676.2116538461539, "end": 676.3716538461539, "probability": 0.998046875 }, { "word": "了", "start": 676.3716538461539, "end": 676.531653846154, "probability": 0.99560546875 }, { "word": "生", "start": 676.531653846154, "end": 676.7116538461539, "probability": 0.99169921875 }, { "word": "成", "start": 676.7116538461539, "end": 676.9716538461539, "probability": 0.9951171875 }, { "word": "解", "start": 676.9716538461539, "end": 677.131653846154, "probability": 0.98681640625 }, { "word": "码", "start": 677.131653846154, "end": 677.3316538461539, "probability": 0.99658203125 }, { "word": "器", "start": 677.3316538461539, "end": 677.4916538461539, "probability": 0.9951171875 }, { "word": "全", "start": 677.4916538461539, "end": 677.771653846154, "probability": 0.99267578125 }, { "word": "局", "start": 677.771653846154, "end": 677.931653846154, "probability": 0.9970703125 }, { "word": "K", "start": 677.931653846154, "end": 678.131653846154, "probability": 0.80517578125 }, { "word": "V", "start": 678.131653846154, "end": 678.3316538461539, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.1042110481072421, "compression_ratio": 1.3841059602649006, "no_speech_prob": 3.177028431999851e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "D组以一个完整层开头后面接三个复用层后面的各组则以一个众索引层开头再接三个复用层所以综合起来看你基本上同时获得了两个层面的成本节省首先CD避免了仅仅为了生成解码器全局KV就让整个预填充跑变模型上半部分相反它直接从最终编码器表示派生这份KV然后借助CSAR一旦某个解码器层生成了这份KV接下来的几层" } ], "id": 327, "review_annotation": "", "row_id": 327, "bounded_row_id": 327, "parent_row_ids": [ 327 ], "parent_row_id": 327, "packet_id": 24, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 661.311653846154, "packet_end": 690.499639423077, "raw_text": "CD避免了仅仅为了生成解码器全局KV", "selected_text": "CD避免了仅仅为了生成解码器全局KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 327 ], "packet_id": 24, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 678.3316538461539, "end": 681.391653846154, "text": "就让整个预填充跑变模型上半部分", "raw_asr_text": "就让整个预填充跑变模型上半部分", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 24, "words": [ { "word": "就", "start": 678.3316538461539, "end": 678.751653846154, "probability": 0.99560546875 }, { "word": "让", "start": 678.751653846154, "end": 678.931653846154, "probability": 0.9990234375 }, { "word": "整", "start": 678.931653846154, "end": 679.171653846154, "probability": 0.99853515625 }, { "word": "个", "start": 679.171653846154, "end": 679.4516538461539, "probability": 0.9990234375 }, { "word": "预", "start": 679.4516538461539, "end": 679.6116538461539, "probability": 0.9111328125 }, { "word": "填", "start": 679.6116538461539, "end": 679.811653846154, "probability": 0.9873046875 }, { "word": "充", "start": 679.811653846154, "end": 680.011653846154, "probability": 0.9970703125 }, { "word": "跑", "start": 680.011653846154, "end": 680.2316538461539, "probability": 0.47900390625 }, { "word": "变", "start": 680.2316538461539, "end": 680.391653846154, "probability": 0.6806640625 }, { "word": "模", "start": 680.391653846154, "end": 680.511653846154, "probability": 0.99853515625 }, { "word": "型", "start": 680.511653846154, "end": 680.7116538461539, "probability": 0.9990234375 }, { "word": "上", "start": 680.7116538461539, "end": 680.9516538461539, "probability": 0.9970703125 }, { "word": "半", "start": 680.9516538461539, "end": 681.0916538461539, "probability": 0.96240234375 }, { "word": "部分", "start": 681.0916538461539, "end": 681.391653846154, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.1042110481072421, "compression_ratio": 1.3841059602649006, "no_speech_prob": 3.177028431999851e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "D组以一个完整层开头后面接三个复用层后面的各组则以一个众索引层开头再接三个复用层所以综合起来看你基本上同时获得了两个层面的成本节省首先CD避免了仅仅为了生成解码器全局KV就让整个预填充跑变模型上半部分相反它直接从最终编码器表示派生这份KV然后借助CSAR一旦某个解码器层生成了这份KV接下来的几层" } ], "id": 328, "review_annotation": "", "row_id": 328, "bounded_row_id": 328, "parent_row_ids": [ 328 ], "parent_row_id": 328, "packet_id": 24, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 661.311653846154, "packet_end": 690.499639423077, "raw_text": "就让整个预填充跑变模型上半部分", "selected_text": "就让整个预填充跑变模型上半部分", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 328 ], "packet_id": 24, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 681.391653846154, "end": 681.9516538461539, "text": "相反", "raw_asr_text": "相反", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 24, "words": [ { "word": "相", "start": 681.391653846154, "end": 681.751653846154, "probability": 0.99755859375 }, { "word": "反", "start": 681.751653846154, "end": 681.9516538461539, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1042110481072421, "compression_ratio": 1.3841059602649006, "no_speech_prob": 3.177028431999851e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "D组以一个完整层开头后面接三个复用层后面的各组则以一个众索引层开头再接三个复用层所以综合起来看你基本上同时获得了两个层面的成本节省首先CD避免了仅仅为了生成解码器全局KV就让整个预填充跑变模型上半部分相反它直接从最终编码器表示派生这份KV然后借助CSAR一旦某个解码器层生成了这份KV接下来的几层" } ], "id": 329, "review_annotation": "", "row_id": 329, "bounded_row_id": 329, "parent_row_ids": [ 329 ], "parent_row_id": 329, "packet_id": 24, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 661.311653846154, "packet_end": 690.499639423077, "raw_text": "相反", "selected_text": "相反", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 329 ], "packet_id": 24, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 681.9516538461539, "end": 684.9716538461539, "text": "它直接从最终编码器表示派生这份KV", "raw_asr_text": "它直接从最终编码器表示派生这份KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 24, "words": [ { "word": "它", "start": 681.9516538461539, "end": 682.1916538461539, "probability": 0.75341796875 }, { "word": "直接", "start": 682.1916538461539, "end": 682.531653846154, "probability": 0.99951171875 }, { "word": "从", "start": 682.531653846154, "end": 682.771653846154, "probability": 0.99951171875 }, { "word": "最", "start": 682.771653846154, "end": 682.931653846154, "probability": 0.99951171875 }, { "word": "终", "start": 682.931653846154, "end": 683.0716538461539, "probability": 0.9990234375 }, { "word": "编", "start": 683.0716538461539, "end": 683.2116538461539, "probability": 0.9853515625 }, { "word": "码", "start": 683.2116538461539, "end": 683.3716538461539, "probability": 1.0 }, { "word": "器", "start": 683.3716538461539, "end": 683.531653846154, "probability": 1.0 }, { "word": "表示", "start": 683.531653846154, "end": 683.911653846154, "probability": 0.99853515625 }, { "word": "派", "start": 683.911653846154, "end": 684.1916538461539, "probability": 0.990234375 }, { "word": "生", "start": 684.1916538461539, "end": 684.3316538461539, "probability": 0.6787109375 }, { "word": "这", "start": 684.3316538461539, "end": 684.4916538461539, "probability": 0.9990234375 }, { "word": "份", "start": 684.4916538461539, "end": 684.5916538461539, "probability": 0.99658203125 }, { "word": "K", "start": 684.5916538461539, "end": 684.751653846154, "probability": 0.99658203125 }, { "word": "V", "start": 684.751653846154, "end": 684.9716538461539, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1042110481072421, "compression_ratio": 1.3841059602649006, "no_speech_prob": 3.177028431999851e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "D组以一个完整层开头后面接三个复用层后面的各组则以一个众索引层开头再接三个复用层所以综合起来看你基本上同时获得了两个层面的成本节省首先CD避免了仅仅为了生成解码器全局KV就让整个预填充跑变模型上半部分相反它直接从最终编码器表示派生这份KV然后借助CSAR一旦某个解码器层生成了这份KV接下来的几层" } ], "id": 330, "review_annotation": "", "row_id": 330, "bounded_row_id": 330, "parent_row_ids": [ 330 ], "parent_row_id": 330, "packet_id": 24, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 661.311653846154, "packet_end": 690.499639423077, "raw_text": "它直接从最终编码器表示派生这份KV", "selected_text": "它直接从最终编码器表示派生这份KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 330 ], "packet_id": 24, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 684.9716538461539, "end": 686.8316538461539, "text": "然后借助CSAR", "raw_asr_text": "然后借助CSAR", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 24, "words": [ { "word": "然后", "start": 684.9716538461539, "end": 685.7116538461539, "probability": 0.99853515625 }, { "word": "借", "start": 685.7116538461539, "end": 686.051653846154, "probability": 0.99560546875 }, { "word": "助", "start": 686.051653846154, "end": 686.2116538461539, "probability": 0.99853515625 }, { "word": "CS", "start": 686.2116538461539, "end": 686.5916538461539, "probability": 0.54443359375 }, { "word": "AR", "start": 686.5916538461539, "end": 686.8316538461539, "probability": 0.9716796875 } ], "asr_metrics": { "avg_logprob": -0.1042110481072421, "compression_ratio": 1.3841059602649006, "no_speech_prob": 3.177028431999851e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "D组以一个完整层开头后面接三个复用层后面的各组则以一个众索引层开头再接三个复用层所以综合起来看你基本上同时获得了两个层面的成本节省首先CD避免了仅仅为了生成解码器全局KV就让整个预填充跑变模型上半部分相反它直接从最终编码器表示派生这份KV然后借助CSAR一旦某个解码器层生成了这份KV接下来的几层" } ], "id": 331, "review_annotation": "", "row_id": 331, "bounded_row_id": 331, "parent_row_ids": [ 331 ], "parent_row_id": 331, "packet_id": 24, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 661.311653846154, "packet_end": 690.499639423077, "raw_text": "然后借助CSAR", "selected_text": "然后借助CSAR", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 331 ], "packet_id": 24, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 686.8316538461539, "end": 689.631653846154, "text": "一旦某个解码器层生成了这份KV", "raw_asr_text": "一旦某个解码器层生成了这份KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 24, "words": [ { "word": "一", "start": 686.8316538461539, "end": 687.131653846154, "probability": 0.99365234375 }, { "word": "旦", "start": 687.131653846154, "end": 687.3316538461539, "probability": 0.9951171875 }, { "word": "某", "start": 687.3316538461539, "end": 687.5916538461539, "probability": 0.9990234375 }, { "word": "个", "start": 687.5916538461539, "end": 687.7116538461539, "probability": 0.998046875 }, { "word": "解", "start": 687.7116538461539, "end": 687.8516538461539, "probability": 0.99853515625 }, { "word": "码", "start": 687.8516538461539, "end": 688.011653846154, "probability": 1.0 }, { "word": "器", "start": 688.011653846154, "end": 688.171653846154, "probability": 0.99951171875 }, { "word": "层", "start": 688.171653846154, "end": 688.3716538461539, "probability": 0.9677734375 }, { "word": "生", "start": 688.3716538461539, "end": 688.6116538461539, "probability": 0.9853515625 }, { "word": "成", "start": 688.6116538461539, "end": 688.8316538461539, "probability": 0.99951171875 }, { "word": "了", "start": 688.8316538461539, "end": 688.9516538461539, "probability": 0.99609375 }, { "word": "这", "start": 688.9516538461539, "end": 689.0916538461539, "probability": 0.99951171875 }, { "word": "份", "start": 689.0916538461539, "end": 689.251653846154, "probability": 0.99951171875 }, { "word": "K", "start": 689.251653846154, "end": 689.431653846154, "probability": 0.9990234375 }, { "word": "V", "start": 689.431653846154, "end": 689.631653846154, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1042110481072421, "compression_ratio": 1.3841059602649006, "no_speech_prob": 3.177028431999851e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "D组以一个完整层开头后面接三个复用层后面的各组则以一个众索引层开头再接三个复用层所以综合起来看你基本上同时获得了两个层面的成本节省首先CD避免了仅仅为了生成解码器全局KV就让整个预填充跑变模型上半部分相反它直接从最终编码器表示派生这份KV然后借助CSAR一旦某个解码器层生成了这份KV接下来的几层" } ], "id": 332, "review_annotation": "", "row_id": 332, "bounded_row_id": 332, "parent_row_ids": [ 332 ], "parent_row_id": 332, "packet_id": 24, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 661.311653846154, "packet_end": 690.499639423077, "raw_text": "一旦某个解码器层生成了这份KV", "selected_text": "一旦某个解码器层生成了这份KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 332 ], "packet_id": 24, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 689.6996394230771, "end": 690.5396394230771, "text": "接下来的几层", "raw_asr_text": "接下来的几层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "接", "start": 689.5996394230772, "end": 689.9196394230771, "probability": 0.72216796875 }, { "word": "下", "start": 689.9196394230771, "end": 690.0396394230771, "probability": 0.99951171875 }, { "word": "来", "start": 690.0396394230771, "end": 690.1396394230771, "probability": 1.0 }, { "word": "的", "start": 690.1396394230771, "end": 690.2196394230772, "probability": 0.99609375 }, { "word": "几", "start": 690.2196394230772, "end": 690.3996394230771, "probability": 0.9921875 }, { "word": "层", "start": 690.3996394230771, "end": 690.5396394230771, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 333, "review_annotation": "", "row_id": 333, "bounded_row_id": 333, "parent_row_ids": [ 333 ], "parent_row_id": 333, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "接下来的几层", "selected_text": "接下来的几层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 333 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 690.5396394230771, "end": 691.9996394230772, "text": "也不一定需要再生成", "raw_asr_text": "也不一定需要再生成", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "也", "start": 690.5396394230771, "end": 690.6996394230771, "probability": 0.9921875 }, { "word": "不", "start": 690.6996394230771, "end": 690.8796394230772, "probability": 0.998046875 }, { "word": "一定", "start": 690.8796394230772, "end": 691.1196394230772, "probability": 0.9912109375 }, { "word": "需要", "start": 691.1196394230772, "end": 691.3796394230772, "probability": 0.98388671875 }, { "word": "再", "start": 691.3796394230772, "end": 691.6196394230772, "probability": 0.98876953125 }, { "word": "生", "start": 691.6196394230772, "end": 691.8196394230771, "probability": 0.99658203125 }, { "word": "成", "start": 691.8196394230771, "end": 691.9996394230772, "probability": 0.9912109375 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 334, "review_annotation": "", "row_id": 334, "bounded_row_id": 334, "parent_row_ids": [ 334 ], "parent_row_id": 334, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "也不一定需要再生成", "selected_text": "也不一定需要再生成", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 334 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 691.9996394230772, "end": 692.7396394230772, "text": "另一份副本", "raw_asr_text": "另一份副本", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "另", "start": 691.9996394230772, "end": 692.1796394230771, "probability": 0.97998046875 }, { "word": "一", "start": 692.1796394230771, "end": 692.2996394230771, "probability": 0.99951171875 }, { "word": "份", "start": 692.2996394230771, "end": 692.4396394230771, "probability": 0.99853515625 }, { "word": "副", "start": 692.4396394230771, "end": 692.5596394230771, "probability": 0.998046875 }, { "word": "本", "start": 692.5596394230771, "end": 692.7396394230772, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 335, "review_annotation": "", "row_id": 335, "bounded_row_id": 335, "parent_row_ids": [ 335 ], "parent_row_id": 335, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "另一份副本", "selected_text": "另一份副本", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 335 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 692.7396394230772, "end": 693.7796394230771, "text": "而这种组合", "raw_asr_text": "而这种组合", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "而", "start": 692.7396394230772, "end": 693.1396394230771, "probability": 0.97998046875 }, { "word": "这", "start": 693.1396394230771, "end": 693.2996394230771, "probability": 0.998046875 }, { "word": "种", "start": 693.2996394230771, "end": 693.4196394230771, "probability": 0.99853515625 }, { "word": "组", "start": 693.4196394230771, "end": 693.5796394230771, "probability": 0.99951171875 }, { "word": "合", "start": 693.5796394230771, "end": 693.7796394230771, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 336, "review_annotation": "", "row_id": 336, "bounded_row_id": 336, "parent_row_ids": [ 336 ], "parent_row_id": 336, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "而这种组合", "selected_text": "而这种组合", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 336 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 693.7796394230771, "end": 695.5596394230771, "text": "正是DeepSeq的方案激进的多", "raw_asr_text": "正是DeepSeq的方案激进的多", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "正", "start": 693.7796394230771, "end": 693.9796394230772, "probability": 0.9912109375 }, { "word": "是", "start": 693.9796394230772, "end": 694.1396394230771, "probability": 0.99853515625 }, { "word": "De", "start": 694.1396394230771, "end": 694.2796394230771, "probability": 0.82958984375 }, { "word": "ep", "start": 694.2796394230771, "end": 694.3596394230772, "probability": 0.94580078125 }, { "word": "Se", "start": 694.3596394230772, "end": 694.4796394230772, "probability": 0.432861328125 }, { "word": "q", "start": 694.4796394230772, "end": 694.5396394230771, "probability": 0.54345703125 }, { "word": "的", "start": 694.5396394230771, "end": 694.6596394230771, "probability": 0.9443359375 }, { "word": "方", "start": 694.6596394230771, "end": 694.8396394230772, "probability": 0.99951171875 }, { "word": "案", "start": 694.8396394230772, "end": 694.9996394230772, "probability": 0.99853515625 }, { "word": "激", "start": 694.9996394230772, "end": 695.1396394230771, "probability": 0.83203125 }, { "word": "进", "start": 695.1396394230771, "end": 695.2596394230771, "probability": 0.99658203125 }, { "word": "的", "start": 695.2596394230771, "end": 695.3396394230772, "probability": 0.88037109375 }, { "word": "多", "start": 695.3396394230772, "end": 695.5596394230771, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 337, "review_annotation": "", "row_id": 337, "bounded_row_id": 337, "parent_row_ids": [ 337 ], "parent_row_id": 337, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "正是DeepSeq的方案激进的多", "selected_text": "正是DeepSeq的方案激进的多", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 337 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 695.5596394230771, "end": 697.8196394230771, "text": "且专门针对V4.1优化的原因", "raw_asr_text": "且专门针对V4.1优化的原因", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "且", "start": 695.5596394230771, "end": 695.9396394230771, "probability": 0.99365234375 }, { "word": "专", "start": 695.9396394230771, "end": 696.1396394230771, "probability": 1.0 }, { "word": "门", "start": 696.1396394230771, "end": 696.2796394230771, "probability": 0.99951171875 }, { "word": "针", "start": 696.2796394230771, "end": 696.4596394230771, "probability": 0.9990234375 }, { "word": "对", "start": 696.4596394230771, "end": 696.5796394230771, "probability": 0.99951171875 }, { "word": "V", "start": 696.5796394230771, "end": 696.7596394230771, "probability": 0.88134765625 }, { "word": "4.", "start": 696.7596394230771, "end": 696.8996394230771, "probability": 0.994140625 }, { "word": "1", "start": 697.0196394230771, "end": 697.1596394230771, "probability": 0.99462890625 }, { "word": "优", "start": 697.1596394230771, "end": 697.2796394230771, "probability": 0.99853515625 }, { "word": "化", "start": 697.2796394230771, "end": 697.4196394230771, "probability": 0.9990234375 }, { "word": "的", "start": 697.4196394230771, "end": 697.5196394230771, "probability": 0.998046875 }, { "word": "原因", "start": 697.5196394230771, "end": 697.8196394230771, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 338, "review_annotation": "", "row_id": 338, "bounded_row_id": 338, "parent_row_ids": [ 338 ], "parent_row_id": 338, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "且专门针对V4.1优化的原因", "selected_text": "且专门针对V4.1优化的原因", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 338 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 697.8196394230771, "end": 698.6396394230771, "text": "除此之外", "raw_asr_text": "除此之外", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "除", "start": 697.8196394230771, "end": 698.2396394230772, "probability": 0.99462890625 }, { "word": "此", "start": 698.2396394230772, "end": 698.3796394230772, "probability": 1.0 }, { "word": "之", "start": 698.3796394230772, "end": 698.4796394230772, "probability": 1.0 }, { "word": "外", "start": 698.4796394230772, "end": 698.6396394230771, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 339, "review_annotation": "", "row_id": 339, "bounded_row_id": 339, "parent_row_ids": [ 339 ], "parent_row_id": 339, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "除此之外", "selected_text": "除此之外", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 339 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 698.6396394230771, "end": 700.0996394230772, "text": "DeepSeq V4.1的情况", "raw_asr_text": "DeepSeq V4.1的情况", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "De", "start": 698.6396394230771, "end": 698.8596394230772, "probability": 0.9794921875 }, { "word": "ep", "start": 698.8596394230772, "end": 698.9396394230771, "probability": 0.99951171875 }, { "word": "Se", "start": 698.9396394230771, "end": 699.0596394230771, "probability": 0.9931640625 }, { "word": "q", "start": 699.0596394230771, "end": 699.1796394230771, "probability": 0.998046875 }, { "word": " V", "start": 699.1796394230771, "end": 699.3596394230772, "probability": 0.74755859375 }, { "word": "4.", "start": 699.3596394230772, "end": 699.5196394230771, "probability": 0.99853515625 }, { "word": "1", "start": 699.6396394230771, "end": 699.7596394230771, "probability": 0.99951171875 }, { "word": "的", "start": 699.7596394230771, "end": 699.8396394230772, "probability": 0.98828125 }, { "word": "情况", "start": 699.8396394230772, "end": 700.0996394230772, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 340, "review_annotation": "", "row_id": 340, "bounded_row_id": 340, "parent_row_ids": [ 340 ], "parent_row_id": 340, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "DeepSeq V4.1的情况", "selected_text": "DeepSeq V4.1的情况", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 340 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 700.0996394230772, "end": 701.0396394230771, "text": "也要复杂的多", "raw_asr_text": "也要复杂的多", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "也", "start": 700.0996394230772, "end": 700.3796394230772, "probability": 0.9970703125 }, { "word": "要", "start": 700.3796394230772, "end": 700.5396394230771, "probability": 0.99951171875 }, { "word": "复", "start": 700.5396394230771, "end": 700.6596394230771, "probability": 0.9990234375 }, { "word": "杂", "start": 700.6596394230771, "end": 700.7396394230772, "probability": 0.9951171875 }, { "word": "的", "start": 700.7396394230772, "end": 700.8396394230772, "probability": 0.83203125 }, { "word": "多", "start": 700.8396394230772, "end": 701.0396394230771, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 341, "review_annotation": "", "row_id": 341, "bounded_row_id": 341, "parent_row_ids": [ 341 ], "parent_row_id": 341, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "也要复杂的多", "selected_text": "也要复杂的多", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 341 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 701.0396394230771, "end": 702.3396394230772, "text": "它在每一层中", "raw_asr_text": "它在每一层中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "它", "start": 701.0396394230771, "end": 701.6396394230771, "probability": 0.98876953125 }, { "word": "在", "start": 701.6396394230771, "end": 701.7596394230771, "probability": 0.99951171875 }, { "word": "每", "start": 701.7596394230771, "end": 701.9396394230771, "probability": 0.99951171875 }, { "word": "一", "start": 701.9396394230771, "end": 702.0396394230771, "probability": 0.96533203125 }, { "word": "层", "start": 702.0396394230771, "end": 702.1796394230771, "probability": 1.0 }, { "word": "中", "start": 702.1796394230771, "end": 702.3396394230772, "probability": 0.9931640625 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 342, "review_annotation": "", "row_id": 342, "bounded_row_id": 342, "parent_row_ids": [ 342 ], "parent_row_id": 342, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "它在每一层中", "selected_text": "它在每一层中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 342 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 702.3396394230772, "end": 703.3396394230772, "text": "都仍然保留着", "raw_asr_text": "都仍然保留着", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "都", "start": 702.3396394230772, "end": 702.5796394230771, "probability": 0.9951171875 }, { "word": "仍", "start": 702.5796394230771, "end": 702.7596394230771, "probability": 1.0 }, { "word": "然", "start": 702.7596394230771, "end": 702.9196394230771, "probability": 0.99755859375 }, { "word": "保", "start": 702.9196394230771, "end": 703.0796394230771, "probability": 0.99951171875 }, { "word": "留", "start": 703.0796394230771, "end": 703.2196394230772, "probability": 0.99951171875 }, { "word": "着", "start": 703.2196394230772, "end": 703.3396394230772, "probability": 0.99560546875 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 343, "review_annotation": "", "row_id": 343, "bounded_row_id": 343, "parent_row_ids": [ 343 ], "parent_row_id": 343, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "都仍然保留着", "selected_text": "都仍然保留着", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 343 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 703.3396394230772, "end": 704.5796394230771, "text": "滑动窗口注意力分支", "raw_asr_text": "滑动窗口注意力分支", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "滑", "start": 703.3396394230772, "end": 703.5796394230771, "probability": 0.9951171875 }, { "word": "动", "start": 703.5796394230771, "end": 703.6796394230771, "probability": 0.9951171875 }, { "word": "窗", "start": 703.6796394230771, "end": 703.8196394230771, "probability": 0.9970703125 }, { "word": "口", "start": 703.8196394230771, "end": 703.9796394230772, "probability": 0.99951171875 }, { "word": "注意", "start": 703.9796394230772, "end": 704.1596394230771, "probability": 0.98828125 }, { "word": "力", "start": 704.1596394230771, "end": 704.2996394230771, "probability": 0.9990234375 }, { "word": "分", "start": 704.2996394230771, "end": 704.4396394230771, "probability": 0.9990234375 }, { "word": "支", "start": 704.4396394230771, "end": 704.5796394230771, "probability": 0.99072265625 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 344, "review_annotation": "", "row_id": 344, "bounded_row_id": 344, "parent_row_ids": [ 344 ], "parent_row_id": 344, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "滑动窗口注意力分支", "selected_text": "滑动窗口注意力分支", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 344 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 704.5796394230771, "end": 705.8796394230772, "text": "所以每个解码器层", "raw_asr_text": "所以每个解码器层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "所以", "start": 704.5796394230771, "end": 704.9396394230771, "probability": 0.99462890625 }, { "word": "每", "start": 704.9396394230771, "end": 705.1196394230772, "probability": 0.99853515625 }, { "word": "个", "start": 705.1196394230772, "end": 705.2396394230772, "probability": 0.9990234375 }, { "word": "解", "start": 705.2396394230772, "end": 705.3596394230772, "probability": 0.9697265625 }, { "word": "码", "start": 705.3596394230772, "end": 705.5196394230771, "probability": 0.9990234375 }, { "word": "器", "start": 705.5196394230771, "end": 705.6596394230771, "probability": 0.98388671875 }, { "word": "层", "start": 705.6596394230771, "end": 705.8796394230772, "probability": 0.9970703125 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 345, "review_annotation": "", "row_id": 345, "bounded_row_id": 345, "parent_row_ids": [ 345 ], "parent_row_id": 345, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "所以每个解码器层", "selected_text": "所以每个解码器层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 345 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 705.8796394230772, "end": 707.4796394230772, "text": "其实有两种不同的记忆", "raw_asr_text": "其实有两种不同的记忆", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "其实", "start": 705.8796394230772, "end": 706.1396394230771, "probability": 0.99755859375 }, { "word": "有", "start": 706.1396394230771, "end": 706.3396394230772, "probability": 0.99951171875 }, { "word": "两", "start": 706.3396394230772, "end": 706.4996394230772, "probability": 0.998046875 }, { "word": "种", "start": 706.4996394230772, "end": 706.7196394230772, "probability": 1.0 }, { "word": "不同", "start": 706.7196394230772, "end": 706.9396394230771, "probability": 0.99951171875 }, { "word": "的", "start": 706.9396394230771, "end": 707.1196394230772, "probability": 0.99951171875 }, { "word": "记", "start": 707.1196394230772, "end": 707.2796394230771, "probability": 0.87109375 }, { "word": "忆", "start": 707.2796394230771, "end": 707.4796394230772, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 346, "review_annotation": "", "row_id": 346, "bounded_row_id": 346, "parent_row_ids": [ 346 ], "parent_row_id": 346, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "其实有两种不同的记忆", "selected_text": "其实有两种不同的记忆", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 346 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 707.4796394230772, "end": 709.6996394230771, "text": "一种是来自编码器表示", "raw_asr_text": "一种是来自编码器表示", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "一", "start": 707.4796394230772, "end": 707.8396394230772, "probability": 0.99609375 }, { "word": "种", "start": 707.8396394230772, "end": 707.9996394230772, "probability": 0.99951171875 }, { "word": "是", "start": 707.9996394230772, "end": 708.1996394230771, "probability": 0.99951171875 }, { "word": "来", "start": 708.1996394230771, "end": 708.4196394230771, "probability": 0.99853515625 }, { "word": "自", "start": 708.4196394230771, "end": 708.6196394230772, "probability": 1.0 }, { "word": "编", "start": 708.6196394230772, "end": 708.9396394230771, "probability": 0.9921875 }, { "word": "码", "start": 708.9396394230771, "end": 709.1196394230772, "probability": 1.0 }, { "word": "器", "start": 709.1196394230772, "end": 709.3196394230771, "probability": 0.99951171875 }, { "word": "表示", "start": 709.3196394230771, "end": 709.6996394230771, "probability": 0.99658203125 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 347, "review_annotation": "", "row_id": 347, "bounded_row_id": 347, "parent_row_ids": [ 347 ], "parent_row_id": 347, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "一种是来自编码器表示", "selected_text": "一种是来自编码器表示", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 347 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 709.6996394230771, "end": 712.4796394230772, "text": "并通过CSAR共享的全局记忆", "raw_asr_text": "并通过CSAR共享的全局记忆", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "并", "start": 709.6996394230771, "end": 710.0796394230771, "probability": 0.99853515625 }, { "word": "通", "start": 710.0796394230771, "end": 710.2196394230772, "probability": 0.998046875 }, { "word": "过", "start": 710.2196394230772, "end": 710.4196394230771, "probability": 0.99951171875 }, { "word": "CS", "start": 710.4196394230771, "end": 710.7996394230771, "probability": 0.7685546875 }, { "word": "AR", "start": 710.7996394230771, "end": 711.3196394230771, "probability": 0.9912109375 }, { "word": "共", "start": 711.3196394230771, "end": 711.6196394230772, "probability": 0.99609375 }, { "word": "享", "start": 711.6196394230772, "end": 711.7596394230771, "probability": 0.99853515625 }, { "word": "的", "start": 711.7596394230771, "end": 711.8596394230772, "probability": 0.99755859375 }, { "word": "全", "start": 711.8596394230772, "end": 711.9996394230772, "probability": 0.99462890625 }, { "word": "局", "start": 711.9996394230772, "end": 712.1196394230772, "probability": 0.9599609375 }, { "word": "记", "start": 712.1196394230772, "end": 712.2996394230771, "probability": 0.998046875 }, { "word": "忆", "start": 712.2996394230771, "end": 712.4796394230772, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 348, "review_annotation": "", "row_id": 348, "bounded_row_id": 348, "parent_row_ids": [ 348 ], "parent_row_id": 348, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "并通过CSAR共享的全局记忆", "selected_text": "并通过CSAR共享的全局记忆", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 348 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 712.4796394230772, "end": 713.6796394230771, "text": "另一种是来自", "raw_asr_text": "另一种是来自", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "另", "start": 712.4796394230772, "end": 712.7396394230772, "probability": 0.994140625 }, { "word": "一", "start": 712.7396394230772, "end": 712.9196394230771, "probability": 1.0 }, { "word": "种", "start": 712.9196394230771, "end": 713.0796394230771, "probability": 0.99951171875 }, { "word": "是", "start": 713.0796394230771, "end": 713.3196394230771, "probability": 0.9990234375 }, { "word": "来", "start": 713.3196394230771, "end": 713.5196394230771, "probability": 0.99658203125 }, { "word": "自", "start": 713.5196394230771, "end": 713.6796394230771, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 349, "review_annotation": "", "row_id": 349, "bounded_row_id": 349, "parent_row_ids": [ 349 ], "parent_row_id": 349, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "另一种是来自", "selected_text": "另一种是来自", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 349 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 713.6796394230771, "end": 715.0596394230771, "text": "该解码器层自身", "raw_asr_text": "该解码器层自身", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "该", "start": 713.6796394230771, "end": 713.9196394230771, "probability": 0.9951171875 }, { "word": "解", "start": 713.9196394230771, "end": 714.0796394230771, "probability": 0.99951171875 }, { "word": "码", "start": 714.0796394230771, "end": 714.2396394230772, "probability": 1.0 }, { "word": "器", "start": 714.2396394230772, "end": 714.4396394230771, "probability": 0.99951171875 }, { "word": "层", "start": 714.4396394230771, "end": 714.6396394230771, "probability": 0.99853515625 }, { "word": "自", "start": 714.6396394230771, "end": 714.8396394230772, "probability": 0.99853515625 }, { "word": "身", "start": 714.8396394230772, "end": 715.0596394230771, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 350, "review_annotation": "", "row_id": 350, "bounded_row_id": 350, "parent_row_ids": [ 350 ], "parent_row_id": 350, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "该解码器层自身", "selected_text": "该解码器层自身", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 350 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 715.0596394230771, "end": 717.1196394230772, "text": "隐藏状态的局部进气记忆", "raw_asr_text": "隐藏状态的局部进气记忆", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 25, "words": [ { "word": "隐", "start": 715.0596394230771, "end": 715.2796394230771, "probability": 0.998046875 }, { "word": "藏", "start": 715.2796394230771, "end": 715.4396394230771, "probability": 1.0 }, { "word": "状", "start": 715.4396394230771, "end": 715.6396394230771, "probability": 0.9892578125 }, { "word": "态", "start": 715.6396394230771, "end": 715.7996394230771, "probability": 1.0 }, { "word": "的", "start": 715.7996394230771, "end": 715.9396394230771, "probability": 0.99951171875 }, { "word": "局", "start": 715.9396394230771, "end": 716.1396394230771, "probability": 0.99951171875 }, { "word": "部", "start": 716.1396394230771, "end": 716.3596394230772, "probability": 0.99755859375 }, { "word": "进", "start": 716.3596394230772, "end": 716.5396394230771, "probability": 0.91162109375 }, { "word": "气", "start": 716.5396394230771, "end": 716.7396394230772, "probability": 0.8125 }, { "word": "记", "start": 716.7396394230772, "end": 716.9596394230771, "probability": 0.99755859375 }, { "word": "忆", "start": 716.9596394230771, "end": 717.1196394230772, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08804100036621093, "compression_ratio": 1.3059490084985836, "no_speech_prob": 4.230544387029411e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这份KV接下来的几层也不一定需要再生成另一份副本而这种组合正是DeepSeq的方案激进的多且专门针对V4.1优化的原因除此之外DeepSeq V4.1的情况也要复杂的多它在每一层中都仍然保留着滑动窗口注意力分支所以每个解码器层其实有两种不同的记忆一种是来自编码器表示并通过CSAR共享的全局记忆另一种是来自该解码器层自身隐藏状态的局部进气记忆所以这里有一点" } ], "id": 351, "review_annotation": "", "row_id": 351, "bounded_row_id": 351, "parent_row_ids": [ 351 ], "parent_row_id": 351, "packet_id": 25, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 688.8996394230771, "packet_end": 718.087625, "raw_text": "隐藏状态的局部进气记忆", "selected_text": "隐藏状态的局部进气记忆", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 351 ], "packet_id": 25, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 717.287625, "end": 718.9676250000001, "text": "所以这里有一点很需要澄清", "raw_asr_text": "所以这里有一点很需要澄清", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 26, "words": [ { "word": "所以", "start": 717.047625, "end": 717.407625, "probability": 0.196044921875 }, { "word": "这里", "start": 717.407625, "end": 717.7076250000001, "probability": 0.96337890625 }, { "word": "有一", "start": 717.7076250000001, "end": 717.9476250000001, "probability": 0.962890625 }, { "word": "点", "start": 717.9476250000001, "end": 718.0876250000001, "probability": 0.99951171875 }, { "word": "很", "start": 718.0876250000001, "end": 718.2676250000001, "probability": 0.99169921875 }, { "word": "需要", "start": 718.2676250000001, "end": 718.5276250000001, "probability": 0.99853515625 }, { "word": "澄", "start": 718.5276250000001, "end": 718.7676250000001, "probability": 0.998046875 }, { "word": "清", "start": 718.7676250000001, "end": 718.9676250000001, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.10055430562872636, "compression_ratio": 1.328173374613003, "no_speech_prob": 6.34588354087029e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所以这里有一点很需要澄清CSAR的KV共享只适用于全局注意力分支这并不意味着滑动窗口注意力的KV缓存会在各层之间共享每个层仍然保留自己的局部SWA KV缓存因为这种局部记忆是由该层实际的隐藏状态产生的因此即使5个CSAR曾共享同一个全局KV它们仍然可以有5份不同的滑动窗口KV缓存除此之外SWA只保留最近的128个TOKENS因此全局" } ], "id": 352, "review_annotation": "", "row_id": 352, "bounded_row_id": 352, "parent_row_ids": [ 352 ], "parent_row_id": 352, "packet_id": 26, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 716.4876250000001, "packet_end": 745.6756105769231, "raw_text": "所以这里有一点很需要澄清", "selected_text": "所以这里有一点很需要澄清", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 352 ], "packet_id": 26, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 718.9676250000001, "end": 722.3476250000001, "text": "CSAR的KV共享只适用于全局注意力分支", "raw_asr_text": "CSAR的KV共享只适用于全局注意力分支", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 26, "words": [ { "word": "CS", "start": 718.9676250000001, "end": 719.3876250000001, "probability": 0.72021484375 }, { "word": "AR", "start": 719.3876250000001, "end": 719.6876250000001, "probability": 0.94091796875 }, { "word": "的", "start": 719.6876250000001, "end": 719.8876250000001, "probability": 0.99169921875 }, { "word": "K", "start": 719.8876250000001, "end": 720.1076250000001, "probability": 0.94384765625 }, { "word": "V", "start": 720.1076250000001, "end": 720.2676250000001, "probability": 0.986328125 }, { "word": "共", "start": 720.2676250000001, "end": 720.4276250000001, "probability": 0.99462890625 }, { "word": "享", "start": 720.4276250000001, "end": 720.6476250000001, "probability": 0.990234375 }, { "word": "只", "start": 720.6476250000001, "end": 720.8476250000001, "probability": 0.97314453125 }, { "word": "适", "start": 720.8476250000001, "end": 720.9676250000001, "probability": 0.998046875 }, { "word": "用", "start": 720.9676250000001, "end": 721.1076250000001, "probability": 0.998046875 }, { "word": "于", "start": 721.1076250000001, "end": 721.2476250000001, "probability": 0.99462890625 }, { "word": "全", "start": 721.2476250000001, "end": 721.5076250000001, "probability": 0.98974609375 }, { "word": "局", "start": 721.5076250000001, "end": 721.667625, "probability": 0.99853515625 }, { "word": "注意", "start": 721.667625, "end": 721.8676250000001, "probability": 0.99365234375 }, { "word": "力", "start": 721.8676250000001, "end": 722.0276250000001, "probability": 0.99951171875 }, { "word": "分", "start": 722.0276250000001, "end": 722.1876250000001, "probability": 0.99951171875 }, { "word": "支", "start": 722.1876250000001, "end": 722.3476250000001, "probability": 0.9931640625 } ], "asr_metrics": { "avg_logprob": -0.10055430562872636, "compression_ratio": 1.328173374613003, "no_speech_prob": 6.34588354087029e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所以这里有一点很需要澄清CSAR的KV共享只适用于全局注意力分支这并不意味着滑动窗口注意力的KV缓存会在各层之间共享每个层仍然保留自己的局部SWA KV缓存因为这种局部记忆是由该层实际的隐藏状态产生的因此即使5个CSAR曾共享同一个全局KV它们仍然可以有5份不同的滑动窗口KV缓存除此之外SWA只保留最近的128个TOKENS因此全局" } ], "id": 353, "review_annotation": "", "row_id": 353, "bounded_row_id": 353, "parent_row_ids": [ 353 ], "parent_row_id": 353, "packet_id": 26, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 716.4876250000001, "packet_end": 745.6756105769231, "raw_text": "CSAR的KV共享只适用于全局注意力分支", "selected_text": "CSAR的KV共享只适用于全局注意力分支", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 353 ], "packet_id": 26, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 722.3476250000001, "end": 725.5076250000001, "text": "这并不意味着滑动窗口注意力的KV缓存", "raw_asr_text": "这并不意味着滑动窗口注意力的KV缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 26, "words": [ { "word": "这", "start": 722.3476250000001, "end": 722.9276250000001, "probability": 0.9736328125 }, { "word": "并", "start": 722.9276250000001, "end": 723.1276250000001, "probability": 0.99951171875 }, { "word": "不", "start": 723.1276250000001, "end": 723.2676250000001, "probability": 0.99853515625 }, { "word": "意", "start": 723.2676250000001, "end": 723.407625, "probability": 0.9990234375 }, { "word": "味", "start": 723.407625, "end": 723.5076250000001, "probability": 0.9970703125 }, { "word": "着", "start": 723.5076250000001, "end": 723.6276250000001, "probability": 0.9951171875 }, { "word": "滑", "start": 723.6276250000001, "end": 723.8876250000001, "probability": 0.9619140625 }, { "word": "动", "start": 723.8876250000001, "end": 724.047625, "probability": 0.998046875 }, { "word": "窗", "start": 724.047625, "end": 724.2076250000001, "probability": 0.994140625 }, { "word": "口", "start": 724.2076250000001, "end": 724.4276250000001, "probability": 0.99951171875 }, { "word": "注意", "start": 724.4276250000001, "end": 724.6276250000001, "probability": 0.98779296875 }, { "word": "力", "start": 724.6276250000001, "end": 724.8076250000001, "probability": 1.0 }, { "word": "的", "start": 724.8076250000001, "end": 724.907625, "probability": 0.9990234375 }, { "word": "K", "start": 724.907625, "end": 725.0276250000001, "probability": 0.98876953125 }, { "word": "V", "start": 725.0276250000001, "end": 725.167625, "probability": 0.99951171875 }, { "word": "缓", "start": 725.167625, "end": 725.3276250000001, "probability": 0.99755859375 }, { "word": "存", "start": 725.3276250000001, "end": 725.5076250000001, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10055430562872636, "compression_ratio": 1.328173374613003, "no_speech_prob": 6.34588354087029e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所以这里有一点很需要澄清CSAR的KV共享只适用于全局注意力分支这并不意味着滑动窗口注意力的KV缓存会在各层之间共享每个层仍然保留自己的局部SWA KV缓存因为这种局部记忆是由该层实际的隐藏状态产生的因此即使5个CSAR曾共享同一个全局KV它们仍然可以有5份不同的滑动窗口KV缓存除此之外SWA只保留最近的128个TOKENS因此全局" } ], "id": 354, "review_annotation": "", "row_id": 354, "bounded_row_id": 354, "parent_row_ids": [ 354 ], "parent_row_id": 354, "packet_id": 26, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 716.4876250000001, "packet_end": 745.6756105769231, "raw_text": "这并不意味着滑动窗口注意力的KV缓存", "selected_text": "这并不意味着滑动窗口注意力的KV缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 354 ], "packet_id": 26, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 725.5076250000001, "end": 726.787625, "text": "会在各层之间共享", "raw_asr_text": "会在各层之间共享", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 26, "words": [ { "word": "会", "start": 725.5076250000001, "end": 725.6876250000001, "probability": 0.99609375 }, { "word": "在", "start": 725.6876250000001, "end": 725.8276250000001, "probability": 0.99951171875 }, { "word": "各", "start": 725.8276250000001, "end": 726.0076250000001, "probability": 0.99560546875 }, { "word": "层", "start": 726.0076250000001, "end": 726.1876250000001, "probability": 0.9990234375 }, { "word": "之", "start": 726.1876250000001, "end": 726.3476250000001, "probability": 0.99169921875 }, { "word": "间", "start": 726.3476250000001, "end": 726.4476250000001, "probability": 1.0 }, { "word": "共", "start": 726.4476250000001, "end": 726.5876250000001, "probability": 0.99853515625 }, { "word": "享", "start": 726.5876250000001, "end": 726.787625, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.10055430562872636, "compression_ratio": 1.328173374613003, "no_speech_prob": 6.34588354087029e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所以这里有一点很需要澄清CSAR的KV共享只适用于全局注意力分支这并不意味着滑动窗口注意力的KV缓存会在各层之间共享每个层仍然保留自己的局部SWA KV缓存因为这种局部记忆是由该层实际的隐藏状态产生的因此即使5个CSAR曾共享同一个全局KV它们仍然可以有5份不同的滑动窗口KV缓存除此之外SWA只保留最近的128个TOKENS因此全局" } ], "id": 355, "review_annotation": "", "row_id": 355, "bounded_row_id": 355, "parent_row_ids": [ 355 ], "parent_row_id": 355, "packet_id": 26, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 716.4876250000001, "packet_end": 745.6756105769231, "raw_text": "会在各层之间共享", "selected_text": "会在各层之间共享", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 355 ], "packet_id": 26, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 726.787625, "end": 730.4476250000001, "text": "每个层仍然保留自己的局部SWA KV缓存", "raw_asr_text": "每个层仍然保留自己的局部SWA KV缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 26, "words": [ { "word": "每", "start": 726.787625, "end": 727.1276250000001, "probability": 0.9833984375 }, { "word": "个", "start": 727.1276250000001, "end": 727.2676250000001, "probability": 0.99462890625 }, { "word": "层", "start": 727.2676250000001, "end": 727.4676250000001, "probability": 0.99951171875 }, { "word": "仍", "start": 727.4676250000001, "end": 727.6476250000001, "probability": 0.9990234375 }, { "word": "然", "start": 727.6476250000001, "end": 727.8076250000001, "probability": 0.99951171875 }, { "word": "保", "start": 727.8076250000001, "end": 728.0076250000001, "probability": 0.99951171875 }, { "word": "留", "start": 728.0076250000001, "end": 728.2076250000001, "probability": 0.9990234375 }, { "word": "自己", "start": 728.2076250000001, "end": 728.4876250000001, "probability": 0.99755859375 }, { "word": "的", "start": 728.4876250000001, "end": 728.6476250000001, "probability": 0.99169921875 }, { "word": "局", "start": 728.6476250000001, "end": 728.8476250000001, "probability": 0.99951171875 }, { "word": "部", "start": 728.8476250000001, "end": 729.047625, "probability": 0.9990234375 }, { "word": "S", "start": 729.047625, "end": 729.2676250000001, "probability": 0.99365234375 }, { "word": "WA", "start": 729.2676250000001, "end": 729.6876250000001, "probability": 0.97802734375 }, { "word": " K", "start": 729.6876250000001, "end": 729.8276250000001, "probability": 0.25732421875 }, { "word": "V", "start": 729.8276250000001, "end": 730.0076250000001, "probability": 0.9990234375 }, { "word": "缓", "start": 730.0076250000001, "end": 730.2076250000001, "probability": 0.9990234375 }, { "word": "存", "start": 730.2076250000001, "end": 730.4476250000001, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10055430562872636, "compression_ratio": 1.328173374613003, "no_speech_prob": 6.34588354087029e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所以这里有一点很需要澄清CSAR的KV共享只适用于全局注意力分支这并不意味着滑动窗口注意力的KV缓存会在各层之间共享每个层仍然保留自己的局部SWA KV缓存因为这种局部记忆是由该层实际的隐藏状态产生的因此即使5个CSAR曾共享同一个全局KV它们仍然可以有5份不同的滑动窗口KV缓存除此之外SWA只保留最近的128个TOKENS因此全局" } ], "id": 356, "review_annotation": "", "row_id": 356, "bounded_row_id": 356, "parent_row_ids": [ 356 ], "parent_row_id": 356, "packet_id": 26, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 716.4876250000001, "packet_end": 745.6756105769231, "raw_text": "每个层仍然保留自己的局部SWA KV缓存", "selected_text": "每个层仍然保留自己的局部SWA KV缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 356 ], "packet_id": 26, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 730.4476250000001, "end": 733.9876250000001, "text": "因为这种局部记忆是由该层实际的隐藏状态产生的", "raw_asr_text": "因为这种局部记忆是由该层实际的隐藏状态产生的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 26, "words": [ { "word": "因为", "start": 730.4476250000001, "end": 730.8676250000001, "probability": 0.9853515625 }, { "word": "这", "start": 730.8676250000001, "end": 731.0276250000001, "probability": 0.99853515625 }, { "word": "种", "start": 731.0276250000001, "end": 731.1276250000001, "probability": 0.99951171875 }, { "word": "局", "start": 731.1276250000001, "end": 731.2676250000001, "probability": 1.0 }, { "word": "部", "start": 731.2676250000001, "end": 731.407625, "probability": 0.99951171875 }, { "word": "记", "start": 731.407625, "end": 731.547625, "probability": 0.90478515625 }, { "word": "忆", "start": 731.547625, "end": 731.7076250000001, "probability": 0.9990234375 }, { "word": "是", "start": 731.7076250000001, "end": 731.9476250000001, "probability": 0.99658203125 }, { "word": "由", "start": 731.9476250000001, "end": 732.047625, "probability": 0.9990234375 }, { "word": "该", "start": 732.047625, "end": 732.2076250000001, "probability": 0.9990234375 }, { "word": "层", "start": 732.2076250000001, "end": 732.407625, "probability": 1.0 }, { "word": "实", "start": 732.407625, "end": 732.6476250000001, "probability": 0.99853515625 }, { "word": "际", "start": 732.6476250000001, "end": 732.787625, "probability": 1.0 }, { "word": "的", "start": 732.787625, "end": 732.8876250000001, "probability": 0.99609375 }, { "word": "隐", "start": 732.8876250000001, "end": 733.0876250000001, "probability": 0.998046875 }, { "word": "藏", "start": 733.0876250000001, "end": 733.2476250000001, "probability": 1.0 }, { "word": "状", "start": 733.2476250000001, "end": 733.407625, "probability": 0.998046875 }, { "word": "态", "start": 733.407625, "end": 733.547625, "probability": 1.0 }, { "word": "产", "start": 733.547625, "end": 733.7076250000001, "probability": 0.9990234375 }, { "word": "生", "start": 733.7076250000001, "end": 733.8676250000001, "probability": 0.99951171875 }, { "word": "的", "start": 733.8676250000001, "end": 733.9876250000001, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.10055430562872636, "compression_ratio": 1.328173374613003, "no_speech_prob": 6.34588354087029e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所以这里有一点很需要澄清CSAR的KV共享只适用于全局注意力分支这并不意味着滑动窗口注意力的KV缓存会在各层之间共享每个层仍然保留自己的局部SWA KV缓存因为这种局部记忆是由该层实际的隐藏状态产生的因此即使5个CSAR曾共享同一个全局KV它们仍然可以有5份不同的滑动窗口KV缓存除此之外SWA只保留最近的128个TOKENS因此全局" } ], "id": 357, "review_annotation": "", "row_id": 357, "bounded_row_id": 357, "parent_row_ids": [ 357 ], "parent_row_id": 357, "packet_id": 26, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 716.4876250000001, "packet_end": 745.6756105769231, "raw_text": "因为这种局部记忆是由该层实际的隐藏状态产生的", "selected_text": "因为这种局部记忆是由该层实际的隐藏状态产生的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 357 ], "packet_id": 26, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 733.9876250000001, "end": 737.4676250000001, "text": "因此即使5个CSAR曾共享同一个全局KV", "raw_asr_text": "因此即使5个CSAR曾共享同一个全局KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 26, "words": [ { "word": "因", "start": 733.9876250000001, "end": 734.3876250000001, "probability": 0.99169921875 }, { "word": "此", "start": 734.3876250000001, "end": 734.5876250000001, "probability": 1.0 }, { "word": "即", "start": 734.5876250000001, "end": 734.8076250000001, "probability": 0.97900390625 }, { "word": "使", "start": 734.8076250000001, "end": 735.0076250000001, "probability": 0.99755859375 }, { "word": "5", "start": 735.0076250000001, "end": 735.167625, "probability": 0.73046875 }, { "word": "个", "start": 735.167625, "end": 735.287625, "probability": 0.99560546875 }, { "word": "CS", "start": 735.287625, "end": 735.5276250000001, "probability": 0.9736328125 }, { "word": "AR", "start": 735.5276250000001, "end": 735.787625, "probability": 0.9951171875 }, { "word": "曾", "start": 735.787625, "end": 736.0076250000001, "probability": 0.7109375 }, { "word": "共", "start": 736.0076250000001, "end": 736.2676250000001, "probability": 0.99755859375 }, { "word": "享", "start": 736.2676250000001, "end": 736.4476250000001, "probability": 0.99951171875 }, { "word": "同", "start": 736.4476250000001, "end": 736.6276250000001, "probability": 0.99462890625 }, { "word": "一个", "start": 736.6276250000001, "end": 736.8276250000001, "probability": 0.99169921875 }, { "word": "全", "start": 736.8276250000001, "end": 737.0276250000001, "probability": 0.99658203125 }, { "word": "局", "start": 737.0276250000001, "end": 737.1476250000001, "probability": 1.0 }, { "word": "K", "start": 737.1476250000001, "end": 737.3076250000001, "probability": 0.99267578125 }, { "word": "V", "start": 737.3076250000001, "end": 737.4676250000001, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10055430562872636, "compression_ratio": 1.328173374613003, "no_speech_prob": 6.34588354087029e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所以这里有一点很需要澄清CSAR的KV共享只适用于全局注意力分支这并不意味着滑动窗口注意力的KV缓存会在各层之间共享每个层仍然保留自己的局部SWA KV缓存因为这种局部记忆是由该层实际的隐藏状态产生的因此即使5个CSAR曾共享同一个全局KV它们仍然可以有5份不同的滑动窗口KV缓存除此之外SWA只保留最近的128个TOKENS因此全局" } ], "id": 358, "review_annotation": "", "row_id": 358, "bounded_row_id": 358, "parent_row_ids": [ 358 ], "parent_row_id": 358, "packet_id": 26, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 716.4876250000001, "packet_end": 745.6756105769231, "raw_text": "因此即使5个CSAR曾共享同一个全局KV", "selected_text": "因此即使5个CSAR曾共享同一个全局KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 358 ], "packet_id": 26, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 737.4676250000001, "end": 740.6076250000001, "text": "它们仍然可以有5份不同的滑动窗口KV缓存", "raw_asr_text": "它们仍然可以有5份不同的滑动窗口KV缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 26, "words": [ { "word": "它", "start": 737.4676250000001, "end": 737.7076250000001, "probability": 0.7001953125 }, { "word": "们", "start": 737.7076250000001, "end": 737.8276250000001, "probability": 0.99853515625 }, { "word": "仍", "start": 737.8276250000001, "end": 737.9876250000001, "probability": 1.0 }, { "word": "然", "start": 737.9876250000001, "end": 738.1476250000001, "probability": 0.98486328125 }, { "word": "可以", "start": 738.1476250000001, "end": 738.3676250000001, "probability": 0.99609375 }, { "word": "有", "start": 738.3676250000001, "end": 738.5676250000001, "probability": 0.9931640625 }, { "word": "5", "start": 738.5676250000001, "end": 738.7476250000001, "probability": 0.94482421875 }, { "word": "份", "start": 738.7476250000001, "end": 738.8876250000001, "probability": 0.98828125 }, { "word": "不同", "start": 738.8876250000001, "end": 739.1476250000001, "probability": 0.99658203125 }, { "word": "的", "start": 739.1476250000001, "end": 739.3076250000001, "probability": 0.99755859375 }, { "word": "滑", "start": 739.3076250000001, "end": 739.4676250000001, "probability": 0.998046875 }, { "word": "动", "start": 739.4676250000001, "end": 739.5676250000001, "probability": 0.99853515625 }, { "word": "窗", "start": 739.5676250000001, "end": 739.7276250000001, "probability": 0.9931640625 }, { "word": "口", "start": 739.7276250000001, "end": 739.907625, "probability": 1.0 }, { "word": "K", "start": 739.907625, "end": 740.0676250000001, "probability": 0.98193359375 }, { "word": "V", "start": 740.0676250000001, "end": 740.2076250000001, "probability": 1.0 }, { "word": "缓", "start": 740.2076250000001, "end": 740.3676250000001, "probability": 0.9990234375 }, { "word": "存", "start": 740.3676250000001, "end": 740.6076250000001, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10055430562872636, "compression_ratio": 1.328173374613003, "no_speech_prob": 6.34588354087029e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所以这里有一点很需要澄清CSAR的KV共享只适用于全局注意力分支这并不意味着滑动窗口注意力的KV缓存会在各层之间共享每个层仍然保留自己的局部SWA KV缓存因为这种局部记忆是由该层实际的隐藏状态产生的因此即使5个CSAR曾共享同一个全局KV它们仍然可以有5份不同的滑动窗口KV缓存除此之外SWA只保留最近的128个TOKENS因此全局" } ], "id": 359, "review_annotation": "", "row_id": 359, "bounded_row_id": 359, "parent_row_ids": [ 359 ], "parent_row_id": 359, "packet_id": 26, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 716.4876250000001, "packet_end": 745.6756105769231, "raw_text": "它们仍然可以有5份不同的滑动窗口KV缓存", "selected_text": "它们仍然可以有5份不同的滑动窗口KV缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 359 ], "packet_id": 26, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 740.6076250000001, "end": 741.5876250000001, "text": "除此之外", "raw_asr_text": "除此之外", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 26, "words": [ { "word": "除", "start": 740.6076250000001, "end": 741.1276250000001, "probability": 0.9951171875 }, { "word": "此", "start": 741.1276250000001, "end": 741.287625, "probability": 1.0 }, { "word": "之", "start": 741.287625, "end": 741.4276250000001, "probability": 0.99951171875 }, { "word": "外", "start": 741.4276250000001, "end": 741.5876250000001, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.10055430562872636, "compression_ratio": 1.328173374613003, "no_speech_prob": 6.34588354087029e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所以这里有一点很需要澄清CSAR的KV共享只适用于全局注意力分支这并不意味着滑动窗口注意力的KV缓存会在各层之间共享每个层仍然保留自己的局部SWA KV缓存因为这种局部记忆是由该层实际的隐藏状态产生的因此即使5个CSAR曾共享同一个全局KV它们仍然可以有5份不同的滑动窗口KV缓存除此之外SWA只保留最近的128个TOKENS因此全局" } ], "id": 360, "review_annotation": "", "row_id": 360, "bounded_row_id": 360, "parent_row_ids": [ 360 ], "parent_row_id": 360, "packet_id": 26, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 716.4876250000001, "packet_end": 745.6756105769231, "raw_text": "除此之外", "selected_text": "除此之外", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 360 ], "packet_id": 26, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 741.5876250000001, "end": 744.7476250000001, "text": "SWA只保留最近的128个TOKENS", "raw_asr_text": "SWA只保留最近的128个TOKENS", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 26, "words": [ { "word": "S", "start": 741.5876250000001, "end": 741.8876250000001, "probability": 0.83740234375 }, { "word": "WA", "start": 741.8876250000001, "end": 742.3076250000001, "probability": 0.9990234375 }, { "word": "只", "start": 742.3076250000001, "end": 742.547625, "probability": 0.99169921875 }, { "word": "保", "start": 742.547625, "end": 742.7676250000001, "probability": 0.9990234375 }, { "word": "留", "start": 742.7676250000001, "end": 742.9276250000001, "probability": 0.99951171875 }, { "word": "最近", "start": 742.9276250000001, "end": 743.2276250000001, "probability": 0.978515625 }, { "word": "的", "start": 743.2276250000001, "end": 743.3876250000001, "probability": 0.9892578125 }, { "word": "12", "start": 743.3876250000001, "end": 743.8276250000001, "probability": 0.99755859375 }, { "word": "8", "start": 743.8276250000001, "end": 744.1476250000001, "probability": 0.99951171875 }, { "word": "个", "start": 744.1476250000001, "end": 744.287625, "probability": 0.99658203125 }, { "word": "T", "start": 744.287625, "end": 744.407625, "probability": 0.6767578125 }, { "word": "OK", "start": 744.407625, "end": 744.547625, "probability": 0.497314453125 }, { "word": "ENS", "start": 744.547625, "end": 744.7476250000001, "probability": 0.53369140625 } ], "asr_metrics": { "avg_logprob": -0.10055430562872636, "compression_ratio": 1.328173374613003, "no_speech_prob": 6.34588354087029e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "所以这里有一点很需要澄清CSAR的KV共享只适用于全局注意力分支这并不意味着滑动窗口注意力的KV缓存会在各层之间共享每个层仍然保留自己的局部SWA KV缓存因为这种局部记忆是由该层实际的隐藏状态产生的因此即使5个CSAR曾共享同一个全局KV它们仍然可以有5份不同的滑动窗口KV缓存除此之外SWA只保留最近的128个TOKENS因此全局" } ], "id": 361, "review_annotation": "", "row_id": 361, "bounded_row_id": 361, "parent_row_ids": [ 361 ], "parent_row_id": 361, "packet_id": 26, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 716.4876250000001, "packet_end": 745.6756105769231, "raw_text": "SWA只保留最近的128个TOKENS", "selected_text": "SWA只保留最近的128个TOKENS", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 361 ], "packet_id": 26, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 744.8756105769231, "end": 748.7756105769232, "text": "因此全局KV会随着整个100万Token的上下文增长", "raw_asr_text": "因此全局KV会随着整个100万Token的上下文增长", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 27, "words": [ { "word": "因", "start": 744.6356105769231, "end": 744.9956105769231, "probability": 0.5234375 }, { "word": "此", "start": 744.9956105769231, "end": 745.2356105769231, "probability": 0.99951171875 }, { "word": "全", "start": 745.2356105769231, "end": 745.6556105769232, "probability": 0.8857421875 }, { "word": "局", "start": 745.6556105769232, "end": 745.8356105769232, "probability": 0.9990234375 }, { "word": "K", "start": 745.8356105769232, "end": 746.0156105769232, "probability": 0.67431640625 }, { "word": "V", "start": 746.0156105769232, "end": 746.1756105769232, "probability": 0.98828125 }, { "word": "会", "start": 746.1756105769232, "end": 746.3956105769232, "probability": 0.98291015625 }, { "word": "随", "start": 746.3956105769232, "end": 746.6156105769231, "probability": 0.99951171875 }, { "word": "着", "start": 746.6156105769231, "end": 746.7556105769231, "probability": 0.99365234375 }, { "word": "整", "start": 746.7556105769231, "end": 746.9556105769232, "probability": 0.99609375 }, { "word": "个", "start": 746.9556105769232, "end": 747.1956105769232, "probability": 0.99951171875 }, { "word": "100", "start": 747.1956105769232, "end": 747.4356105769232, "probability": 0.9140625 }, { "word": "万", "start": 747.4356105769232, "end": 747.6156105769231, "probability": 0.99755859375 }, { "word": "T", "start": 747.6156105769231, "end": 747.7756105769232, "probability": 0.7578125 }, { "word": "oken", "start": 747.7756105769232, "end": 747.9156105769232, "probability": 0.97265625 }, { "word": "的", "start": 747.9156105769232, "end": 748.0156105769232, "probability": 0.99462890625 }, { "word": "上", "start": 748.0156105769232, "end": 748.1556105769232, "probability": 0.99951171875 }, { "word": "下", "start": 748.1556105769232, "end": 748.2756105769232, "probability": 0.9990234375 }, { "word": "文", "start": 748.2756105769232, "end": 748.4156105769232, "probability": 0.426513671875 }, { "word": "增", "start": 748.4156105769232, "end": 748.5956105769231, "probability": 1.0 }, { "word": "长", "start": 748.5956105769231, "end": 748.7756105769232, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.0945693588256836, "compression_ratio": 1.2322946175637395, "no_speech_prob": 7.991534517470811e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "因此全局KV会随着整个100万Token的上下文增长而运行时SWA缓存实际上始终受限于那个固定窗口但由于它仍未压缩且未每层独有所以仍然不是零成本尤其是要跨大量绘画持久保存这些状态时那么如果我们不共享SWA KV缓存这不就意味着还得把提示词送过这些解码7层吗而麻烦的地方在于这个128Token的窗口实际上会随着层深不断累积第40层" } ], "id": 362, "review_annotation": "", "row_id": 362, "bounded_row_id": 362, "parent_row_ids": [ 362 ], "parent_row_id": 362, "packet_id": 27, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 744.0756105769232, "packet_end": 773.2635961538462, "raw_text": "因此全局KV会随着整个100万Token的上下文增长", "selected_text": "因此全局KV会随着整个100万Token的上下文增长", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 362 ], "packet_id": 27, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 748.7756105769232, "end": 749.7156105769232, "text": "而运行时", "raw_asr_text": "而运行时", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 27, "words": [ { "word": "而", "start": 748.7756105769232, "end": 749.0356105769232, "probability": 0.97119140625 }, { "word": "运", "start": 749.0356105769232, "end": 749.2556105769231, "probability": 1.0 }, { "word": "行", "start": 749.2556105769231, "end": 749.4556105769232, "probability": 0.99951171875 }, { "word": "时", "start": 749.4556105769232, "end": 749.7156105769232, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.0945693588256836, "compression_ratio": 1.2322946175637395, "no_speech_prob": 7.991534517470811e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "因此全局KV会随着整个100万Token的上下文增长而运行时SWA缓存实际上始终受限于那个固定窗口但由于它仍未压缩且未每层独有所以仍然不是零成本尤其是要跨大量绘画持久保存这些状态时那么如果我们不共享SWA KV缓存这不就意味着还得把提示词送过这些解码7层吗而麻烦的地方在于这个128Token的窗口实际上会随着层深不断累积第40层" } ], "id": 363, "review_annotation": "", "row_id": 363, "bounded_row_id": 363, "parent_row_ids": [ 363 ], "parent_row_id": 363, "packet_id": 27, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 744.0756105769232, "packet_end": 773.2635961538462, "raw_text": "而运行时", "selected_text": "而运行时", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 363 ], "packet_id": 27, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 749.7156105769232, "end": 753.6756105769232, "text": "SWA缓存实际上始终受限于那个固定窗口", "raw_asr_text": "SWA缓存实际上始终受限于那个固定窗口", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 27, "words": [ { "word": "S", "start": 749.7156105769232, "end": 750.0756105769232, "probability": 0.81591796875 }, { "word": "WA", "start": 750.0756105769232, "end": 750.5156105769232, "probability": 0.9892578125 }, { "word": "缓", "start": 750.5156105769232, "end": 750.7756105769232, "probability": 0.998046875 }, { "word": "存", "start": 750.7756105769232, "end": 750.9756105769231, "probability": 1.0 }, { "word": "实", "start": 750.9756105769231, "end": 751.3356105769232, "probability": 0.998046875 }, { "word": "际", "start": 751.3356105769232, "end": 751.4756105769231, "probability": 0.99951171875 }, { "word": "上", "start": 751.4756105769231, "end": 751.6156105769231, "probability": 0.99658203125 }, { "word": "始", "start": 751.6156105769231, "end": 751.8556105769231, "probability": 0.9990234375 }, { "word": "终", "start": 751.8556105769231, "end": 752.0356105769232, "probability": 0.9990234375 }, { "word": "受", "start": 752.0356105769232, "end": 752.2156105769232, "probability": 0.998046875 }, { "word": "限", "start": 752.2156105769232, "end": 752.4356105769232, "probability": 0.98876953125 }, { "word": "于", "start": 752.4356105769232, "end": 752.5956105769231, "probability": 0.99951171875 }, { "word": "那个", "start": 752.5956105769231, "end": 752.7956105769232, "probability": 0.86767578125 }, { "word": "固", "start": 752.7956105769232, "end": 753.0956105769231, "probability": 0.998046875 }, { "word": "定", "start": 753.0956105769231, "end": 753.2556105769231, "probability": 0.99951171875 }, { "word": "窗", "start": 753.2556105769231, "end": 753.4356105769232, "probability": 0.9951171875 }, { "word": "口", "start": 753.4356105769232, "end": 753.6756105769232, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.0945693588256836, "compression_ratio": 1.2322946175637395, "no_speech_prob": 7.991534517470811e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "因此全局KV会随着整个100万Token的上下文增长而运行时SWA缓存实际上始终受限于那个固定窗口但由于它仍未压缩且未每层独有所以仍然不是零成本尤其是要跨大量绘画持久保存这些状态时那么如果我们不共享SWA KV缓存这不就意味着还得把提示词送过这些解码7层吗而麻烦的地方在于这个128Token的窗口实际上会随着层深不断累积第40层" } ], "id": 364, "review_annotation": "", "row_id": 364, "bounded_row_id": 364, "parent_row_ids": [ 364 ], "parent_row_id": 364, "packet_id": 27, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 744.0756105769232, "packet_end": 773.2635961538462, "raw_text": "SWA缓存实际上始终受限于那个固定窗口", "selected_text": "SWA缓存实际上始终受限于那个固定窗口", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 364 ], "packet_id": 27, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 753.6756105769232, "end": 756.1956105769232, "text": "但由于它仍未压缩且未每层独有", "raw_asr_text": "但由于它仍未压缩且未每层独有", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 27, "words": [ { "word": "但", "start": 753.6756105769232, "end": 753.9956105769231, "probability": 0.9970703125 }, { "word": "由", "start": 753.9956105769231, "end": 754.1956105769232, "probability": 0.99951171875 }, { "word": "于", "start": 754.1956105769232, "end": 754.2756105769232, "probability": 0.99951171875 }, { "word": "它", "start": 754.2756105769232, "end": 754.4556105769232, "probability": 0.96875 }, { "word": "仍", "start": 754.4556105769232, "end": 754.6556105769232, "probability": 0.8134765625 }, { "word": "未", "start": 754.6556105769232, "end": 754.7956105769232, "probability": 0.54833984375 }, { "word": "压", "start": 754.7956105769232, "end": 754.9956105769231, "probability": 0.998046875 }, { "word": "缩", "start": 754.9956105769231, "end": 755.1756105769232, "probability": 0.9990234375 }, { "word": "且", "start": 755.1756105769232, "end": 755.3556105769231, "probability": 0.99853515625 }, { "word": "未", "start": 755.3556105769231, "end": 755.5156105769232, "probability": 0.99365234375 }, { "word": "每", "start": 755.5156105769232, "end": 755.7156105769232, "probability": 0.73828125 }, { "word": "层", "start": 755.7156105769232, "end": 755.8756105769231, "probability": 0.998046875 }, { "word": "独", "start": 755.8756105769231, "end": 756.0756105769232, "probability": 0.994140625 }, { "word": "有", "start": 756.0756105769232, "end": 756.1956105769232, "probability": 0.98095703125 } ], "asr_metrics": { "avg_logprob": -0.0945693588256836, "compression_ratio": 1.2322946175637395, "no_speech_prob": 7.991534517470811e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "因此全局KV会随着整个100万Token的上下文增长而运行时SWA缓存实际上始终受限于那个固定窗口但由于它仍未压缩且未每层独有所以仍然不是零成本尤其是要跨大量绘画持久保存这些状态时那么如果我们不共享SWA KV缓存这不就意味着还得把提示词送过这些解码7层吗而麻烦的地方在于这个128Token的窗口实际上会随着层深不断累积第40层" } ], "id": 365, "review_annotation": "", "row_id": 365, "bounded_row_id": 365, "parent_row_ids": [ 365 ], "parent_row_id": 365, "packet_id": 27, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 744.0756105769232, "packet_end": 773.2635961538462, "raw_text": "但由于它仍未压缩且未每层独有", "selected_text": "但由于它仍未压缩且未每层独有", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 365 ], "packet_id": 27, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 756.1956105769232, "end": 757.7556105769231, "text": "所以仍然不是零成本", "raw_asr_text": "所以仍然不是零成本", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 27, "words": [ { "word": "所以", "start": 756.1956105769232, "end": 756.5356105769232, "probability": 0.99658203125 }, { "word": "仍", "start": 756.5356105769232, "end": 756.8356105769232, "probability": 0.9990234375 }, { "word": "然", "start": 756.8356105769232, "end": 756.9756105769231, "probability": 0.998046875 }, { "word": "不是", "start": 756.9756105769231, "end": 757.2556105769231, "probability": 0.99609375 }, { "word": "零", "start": 757.2556105769231, "end": 757.4556105769232, "probability": 0.9912109375 }, { "word": "成", "start": 757.4556105769232, "end": 757.5956105769231, "probability": 0.99951171875 }, { "word": "本", "start": 757.5956105769231, "end": 757.7556105769231, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.0945693588256836, "compression_ratio": 1.2322946175637395, "no_speech_prob": 7.991534517470811e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "因此全局KV会随着整个100万Token的上下文增长而运行时SWA缓存实际上始终受限于那个固定窗口但由于它仍未压缩且未每层独有所以仍然不是零成本尤其是要跨大量绘画持久保存这些状态时那么如果我们不共享SWA KV缓存这不就意味着还得把提示词送过这些解码7层吗而麻烦的地方在于这个128Token的窗口实际上会随着层深不断累积第40层" } ], "id": 366, "review_annotation": "", "row_id": 366, "bounded_row_id": 366, "parent_row_ids": [ 366 ], "parent_row_id": 366, "packet_id": 27, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 744.0756105769232, "packet_end": 773.2635961538462, "raw_text": "所以仍然不是零成本", "selected_text": "所以仍然不是零成本", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 366 ], "packet_id": 27, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 757.7556105769231, "end": 760.8756105769231, "text": "尤其是要跨大量绘画持久保存这些状态时", "raw_asr_text": "尤其是要跨大量绘画持久保存这些状态时", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 27, "words": [ { "word": "尤", "start": 757.7556105769231, "end": 758.0556105769232, "probability": 0.9990234375 }, { "word": "其", "start": 758.0556105769232, "end": 758.1756105769232, "probability": 1.0 }, { "word": "是", "start": 758.1756105769232, "end": 758.3356105769232, "probability": 0.99951171875 }, { "word": "要", "start": 758.3356105769232, "end": 758.4956105769231, "probability": 0.99658203125 }, { "word": "跨", "start": 758.4956105769231, "end": 758.6556105769232, "probability": 0.98974609375 }, { "word": "大", "start": 758.6556105769232, "end": 758.9356105769232, "probability": 0.9990234375 }, { "word": "量", "start": 758.9356105769232, "end": 759.1156105769231, "probability": 0.99951171875 }, { "word": "绘", "start": 759.1156105769231, "end": 759.2756105769232, "probability": 0.9736328125 }, { "word": "画", "start": 759.2756105769232, "end": 759.4356105769232, "probability": 0.919921875 }, { "word": "持", "start": 759.4356105769232, "end": 759.6556105769232, "probability": 0.9990234375 }, { "word": "久", "start": 759.6556105769232, "end": 759.7956105769232, "probability": 0.9990234375 }, { "word": "保", "start": 759.7956105769232, "end": 759.9556105769232, "probability": 0.99951171875 }, { "word": "存", "start": 759.9556105769232, "end": 760.1356105769231, "probability": 0.99951171875 }, { "word": "这", "start": 760.1356105769231, "end": 760.2956105769232, "probability": 0.9912109375 }, { "word": "些", "start": 760.2956105769232, "end": 760.4156105769232, "probability": 0.99560546875 }, { "word": "状", "start": 760.4156105769232, "end": 760.5556105769232, "probability": 1.0 }, { "word": "态", "start": 760.5556105769232, "end": 760.6956105769232, "probability": 0.99951171875 }, { "word": "时", "start": 760.6956105769232, "end": 760.8756105769231, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.0945693588256836, "compression_ratio": 1.2322946175637395, "no_speech_prob": 7.991534517470811e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "因此全局KV会随着整个100万Token的上下文增长而运行时SWA缓存实际上始终受限于那个固定窗口但由于它仍未压缩且未每层独有所以仍然不是零成本尤其是要跨大量绘画持久保存这些状态时那么如果我们不共享SWA KV缓存这不就意味着还得把提示词送过这些解码7层吗而麻烦的地方在于这个128Token的窗口实际上会随着层深不断累积第40层" } ], "id": 367, "review_annotation": "", "row_id": 367, "bounded_row_id": 367, "parent_row_ids": [ 367 ], "parent_row_id": 367, "packet_id": 27, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 744.0756105769232, "packet_end": 773.2635961538462, "raw_text": "尤其是要跨大量绘画持久保存这些状态时", "selected_text": "尤其是要跨大量绘画持久保存这些状态时", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 367 ], "packet_id": 27, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 760.8756105769231, "end": 763.6956105769232, "text": "那么如果我们不共享SWA KV缓存", "raw_asr_text": "那么如果我们不共享SWA KV缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 27, "words": [ { "word": "那么", "start": 760.8756105769231, "end": 761.2556105769231, "probability": 0.9921875 }, { "word": "如果", "start": 761.2556105769231, "end": 761.5956105769231, "probability": 0.99267578125 }, { "word": "我们", "start": 761.5956105769231, "end": 761.8156105769232, "probability": 0.99951171875 }, { "word": "不", "start": 761.8156105769232, "end": 762.0356105769232, "probability": 0.99951171875 }, { "word": "共", "start": 762.0356105769232, "end": 762.1956105769232, "probability": 0.86865234375 }, { "word": "享", "start": 762.1956105769232, "end": 762.4356105769232, "probability": 0.9794921875 }, { "word": "S", "start": 762.4356105769232, "end": 762.7156105769232, "probability": 0.994140625 }, { "word": "WA", "start": 762.7156105769232, "end": 763.0756105769232, "probability": 0.98876953125 }, { "word": " K", "start": 763.0756105769232, "end": 763.2156105769232, "probability": 0.235595703125 }, { "word": "V", "start": 763.2156105769232, "end": 763.3356105769232, "probability": 0.9970703125 }, { "word": "缓", "start": 763.3356105769232, "end": 763.5156105769232, "probability": 0.9931640625 }, { "word": "存", "start": 763.5156105769232, "end": 763.6956105769232, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.0945693588256836, "compression_ratio": 1.2322946175637395, "no_speech_prob": 7.991534517470811e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "因此全局KV会随着整个100万Token的上下文增长而运行时SWA缓存实际上始终受限于那个固定窗口但由于它仍未压缩且未每层独有所以仍然不是零成本尤其是要跨大量绘画持久保存这些状态时那么如果我们不共享SWA KV缓存这不就意味着还得把提示词送过这些解码7层吗而麻烦的地方在于这个128Token的窗口实际上会随着层深不断累积第40层" } ], "id": 368, "review_annotation": "", "row_id": 368, "bounded_row_id": 368, "parent_row_ids": [ 368 ], "parent_row_id": 368, "packet_id": 27, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 744.0756105769232, "packet_end": 773.2635961538462, "raw_text": "那么如果我们不共享SWA KV缓存", "selected_text": "那么如果我们不共享SWA KV缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 368 ], "packet_id": 27, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 763.6956105769232, "end": 766.9356105769232, "text": "这不就意味着还得把提示词送过这些解码7层吗", "raw_asr_text": "这不就意味着还得把提示词送过这些解码7层吗", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 27, "words": [ { "word": "这", "start": 763.6956105769232, "end": 763.9156105769232, "probability": 0.998046875 }, { "word": "不", "start": 763.9156105769232, "end": 763.9956105769231, "probability": 0.9990234375 }, { "word": "就", "start": 763.9956105769231, "end": 764.1156105769231, "probability": 0.99609375 }, { "word": "意", "start": 764.1156105769231, "end": 764.2556105769231, "probability": 0.99853515625 }, { "word": "味", "start": 764.2556105769231, "end": 764.3356105769232, "probability": 0.99853515625 }, { "word": "着", "start": 764.3356105769232, "end": 764.4356105769232, "probability": 0.998046875 }, { "word": "还", "start": 764.4356105769232, "end": 764.6356105769231, "probability": 0.998046875 }, { "word": "得", "start": 764.6356105769231, "end": 764.7556105769231, "probability": 0.99853515625 }, { "word": "把", "start": 764.7556105769231, "end": 764.9156105769232, "probability": 0.998046875 }, { "word": "提", "start": 764.9156105769232, "end": 765.0756105769232, "probability": 0.9970703125 }, { "word": "示", "start": 765.0756105769232, "end": 765.2356105769231, "probability": 0.9970703125 }, { "word": "词", "start": 765.2356105769231, "end": 765.4356105769232, "probability": 0.99169921875 }, { "word": "送", "start": 765.4356105769232, "end": 765.6356105769231, "probability": 0.99609375 }, { "word": "过", "start": 765.6356105769231, "end": 765.8356105769232, "probability": 0.9990234375 }, { "word": "这", "start": 765.8356105769232, "end": 765.9756105769231, "probability": 0.9931640625 }, { "word": "些", "start": 765.9756105769231, "end": 766.0956105769231, "probability": 0.99951171875 }, { "word": "解", "start": 766.0956105769231, "end": 766.2756105769232, "probability": 0.78662109375 }, { "word": "码", "start": 766.2756105769232, "end": 766.4356105769232, "probability": 0.99853515625 }, { "word": "7", "start": 766.4356105769232, "end": 766.5756105769232, "probability": 0.41552734375 }, { "word": "层", "start": 766.5756105769232, "end": 766.7556105769231, "probability": 0.99951171875 }, { "word": "吗", "start": 766.7556105769231, "end": 766.9356105769232, "probability": 0.97314453125 } ], "asr_metrics": { "avg_logprob": -0.0945693588256836, "compression_ratio": 1.2322946175637395, "no_speech_prob": 7.991534517470811e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "因此全局KV会随着整个100万Token的上下文增长而运行时SWA缓存实际上始终受限于那个固定窗口但由于它仍未压缩且未每层独有所以仍然不是零成本尤其是要跨大量绘画持久保存这些状态时那么如果我们不共享SWA KV缓存这不就意味着还得把提示词送过这些解码7层吗而麻烦的地方在于这个128Token的窗口实际上会随着层深不断累积第40层" } ], "id": 369, "review_annotation": "", "row_id": 369, "bounded_row_id": 369, "parent_row_ids": [ 369 ], "parent_row_id": 369, "packet_id": 27, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 744.0756105769232, "packet_end": 773.2635961538462, "raw_text": "这不就意味着还得把提示词送过这些解码7层吗", "selected_text": "这不就意味着还得把提示词送过这些解码7层吗", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 369 ], "packet_id": 27, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 766.9356105769232, "end": 768.2956105769232, "text": "而麻烦的地方在于", "raw_asr_text": "而麻烦的地方在于", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 27, "words": [ { "word": "而", "start": 766.9356105769232, "end": 767.4556105769232, "probability": 0.728515625 }, { "word": "麻", "start": 767.4556105769232, "end": 767.6356105769231, "probability": 0.99951171875 }, { "word": "烦", "start": 767.6356105769231, "end": 767.7156105769232, "probability": 0.998046875 }, { "word": "的", "start": 767.7156105769232, "end": 767.8156105769232, "probability": 0.99951171875 }, { "word": "地方", "start": 767.8156105769232, "end": 768.0156105769232, "probability": 0.99951171875 }, { "word": "在", "start": 768.0156105769232, "end": 768.1756105769232, "probability": 0.99951171875 }, { "word": "于", "start": 768.1756105769232, "end": 768.2956105769232, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.0945693588256836, "compression_ratio": 1.2322946175637395, "no_speech_prob": 7.991534517470811e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "因此全局KV会随着整个100万Token的上下文增长而运行时SWA缓存实际上始终受限于那个固定窗口但由于它仍未压缩且未每层独有所以仍然不是零成本尤其是要跨大量绘画持久保存这些状态时那么如果我们不共享SWA KV缓存这不就意味着还得把提示词送过这些解码7层吗而麻烦的地方在于这个128Token的窗口实际上会随着层深不断累积第40层" } ], "id": 370, "review_annotation": "", "row_id": 370, "bounded_row_id": 370, "parent_row_ids": [ 370 ], "parent_row_id": 370, "packet_id": 27, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 744.0756105769232, "packet_end": 773.2635961538462, "raw_text": "而麻烦的地方在于", "selected_text": "而麻烦的地方在于", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 370 ], "packet_id": 27, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 768.2956105769232, "end": 772.3956105769232, "text": "这个128Token的窗口实际上会随着层深不断累积", "raw_asr_text": "这个128Token的窗口实际上会随着层深不断累积", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 27, "words": [ { "word": "这个", "start": 768.2956105769232, "end": 768.6156105769231, "probability": 0.99609375 }, { "word": "12", "start": 768.6156105769231, "end": 768.8556105769231, "probability": 0.98583984375 }, { "word": "8", "start": 768.8556105769231, "end": 769.0356105769232, "probability": 0.99560546875 }, { "word": "T", "start": 769.0356105769232, "end": 769.1956105769232, "probability": 0.5712890625 }, { "word": "oken", "start": 769.1956105769232, "end": 769.3556105769231, "probability": 0.9091796875 }, { "word": "的", "start": 769.3556105769231, "end": 769.4356105769232, "probability": 0.9970703125 }, { "word": "窗", "start": 769.4356105769232, "end": 769.5556105769232, "probability": 1.0 }, { "word": "口", "start": 769.5556105769232, "end": 769.7556105769231, "probability": 1.0 }, { "word": "实", "start": 769.7556105769231, "end": 769.9756105769231, "probability": 0.9990234375 }, { "word": "际", "start": 769.9756105769231, "end": 770.1156105769231, "probability": 1.0 }, { "word": "上", "start": 770.1156105769231, "end": 770.2756105769232, "probability": 0.9970703125 }, { "word": "会", "start": 770.2756105769232, "end": 770.4956105769231, "probability": 0.998046875 }, { "word": "随", "start": 770.4956105769231, "end": 770.7356105769231, "probability": 1.0 }, { "word": "着", "start": 770.7356105769231, "end": 770.8956105769232, "probability": 0.9970703125 }, { "word": "层", "start": 770.8956105769232, "end": 771.2956105769232, "probability": 0.9990234375 }, { "word": "深", "start": 771.2956105769232, "end": 771.4556105769232, "probability": 0.470703125 }, { "word": "不", "start": 771.4556105769232, "end": 771.7956105769232, "probability": 0.99560546875 }, { "word": "断", "start": 771.7956105769232, "end": 771.9756105769231, "probability": 0.99853515625 }, { "word": "累", "start": 771.9756105769231, "end": 772.0956105769231, "probability": 0.99853515625 }, { "word": "积", "start": 772.0956105769231, "end": 772.3956105769232, "probability": 0.994140625 } ], "asr_metrics": { "avg_logprob": -0.0945693588256836, "compression_ratio": 1.2322946175637395, "no_speech_prob": 7.991534517470811e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "因此全局KV会随着整个100万Token的上下文增长而运行时SWA缓存实际上始终受限于那个固定窗口但由于它仍未压缩且未每层独有所以仍然不是零成本尤其是要跨大量绘画持久保存这些状态时那么如果我们不共享SWA KV缓存这不就意味着还得把提示词送过这些解码7层吗而麻烦的地方在于这个128Token的窗口实际上会随着层深不断累积第40层" } ], "id": 371, "review_annotation": "", "row_id": 371, "bounded_row_id": 371, "parent_row_ids": [ 371 ], "parent_row_id": 371, "packet_id": 27, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 744.0756105769232, "packet_end": 773.2635961538462, "raw_text": "这个128Token的窗口实际上会随着层深不断累积", "selected_text": "这个128Token的窗口实际上会随着层深不断累积", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 371 ], "packet_id": 27, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 772.4635961538462, "end": 775.9635961538462, "text": "第40层的SWA状态依赖第39层", "raw_asr_text": "第40层的SWA状态依赖第39层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 28, "words": [ { "word": "第", "start": 772.3835961538463, "end": 772.7835961538462, "probability": 0.927734375 }, { "word": "40", "start": 772.7835961538462, "end": 773.0235961538463, "probability": 0.93017578125 }, { "word": "层", "start": 773.0235961538463, "end": 773.2235961538462, "probability": 0.98046875 }, { "word": "的", "start": 773.2235961538462, "end": 773.4035961538463, "probability": 0.98974609375 }, { "word": "S", "start": 773.4035961538463, "end": 773.7035961538462, "probability": 0.80810546875 }, { "word": "WA", "start": 773.7035961538462, "end": 774.1635961538462, "probability": 0.9892578125 }, { "word": "状", "start": 774.1635961538462, "end": 774.4035961538463, "probability": 0.97265625 }, { "word": "态", "start": 774.4035961538463, "end": 774.6635961538462, "probability": 0.8603515625 }, { "word": "依", "start": 774.6635961538462, "end": 774.8635961538463, "probability": 0.990234375 }, { "word": "赖", "start": 774.8635961538463, "end": 775.0035961538463, "probability": 0.966796875 }, { "word": "第", "start": 775.0035961538463, "end": 775.2235961538462, "probability": 0.99072265625 }, { "word": "39", "start": 775.2235961538462, "end": 775.6635961538462, "probability": 0.9951171875 }, { "word": "层", "start": 775.6635961538462, "end": 775.9635961538462, "probability": 0.99658203125 } ], "asr_metrics": { "avg_logprob": -0.13071968680933901, "compression_ratio": 1.3051948051948052, "no_speech_prob": 2.576132709930512e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "不断累积第40层的SWA状态依赖第39层第39层本身又依赖更早的局部窗口依此类推所以Deepseek基本上切断了这条依赖链只把最后128个tokens重放给每个解码7层这正是他们所说的SWA有借重放发挥作用的地方所以不是让100万个tokens经过全部40层而是让100万个经过前20层只让128个经过剩下的20层这意味着现在完整的预填" } ], "id": 372, "review_annotation": "", "row_id": 372, "bounded_row_id": 372, "parent_row_ids": [ 372 ], "parent_row_id": 372, "packet_id": 28, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 771.6635961538462, "packet_end": 800.8515817307693, "raw_text": "第40层的SWA状态依赖第39层", "selected_text": "第40层的SWA状态依赖第39层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 372 ], "packet_id": 28, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 775.9635961538462, "end": 778.8635961538463, "text": "第39层本身又依赖更早的局部窗口", "raw_asr_text": "第39层本身又依赖更早的局部窗口", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 28, "words": [ { "word": "第", "start": 775.9635961538462, "end": 776.2035961538462, "probability": 0.9013671875 }, { "word": "39", "start": 776.2035961538462, "end": 776.5435961538462, "probability": 0.998046875 }, { "word": "层", "start": 776.5435961538462, "end": 776.8635961538463, "probability": 0.9990234375 }, { "word": "本", "start": 776.8635961538463, "end": 777.0435961538462, "probability": 0.98828125 }, { "word": "身", "start": 777.0435961538462, "end": 777.2835961538462, "probability": 1.0 }, { "word": "又", "start": 777.2835961538462, "end": 777.4435961538462, "probability": 0.98828125 }, { "word": "依", "start": 777.4435961538462, "end": 777.6235961538463, "probability": 0.99853515625 }, { "word": "赖", "start": 777.6235961538463, "end": 777.7635961538463, "probability": 0.9951171875 }, { "word": "更", "start": 777.7635961538463, "end": 777.9435961538462, "probability": 0.99755859375 }, { "word": "早", "start": 777.9435961538462, "end": 778.1035961538463, "probability": 0.99951171875 }, { "word": "的", "start": 778.1035961538463, "end": 778.2235961538462, "probability": 0.9921875 }, { "word": "局", "start": 778.2235961538462, "end": 778.3435961538462, "probability": 0.99951171875 }, { "word": "部", "start": 778.3435961538462, "end": 778.5035961538463, "probability": 0.99755859375 }, { "word": "窗", "start": 778.5035961538463, "end": 778.6635961538462, "probability": 0.998046875 }, { "word": "口", "start": 778.6635961538462, "end": 778.8635961538463, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.13071968680933901, "compression_ratio": 1.3051948051948052, "no_speech_prob": 2.576132709930512e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "不断累积第40层的SWA状态依赖第39层第39层本身又依赖更早的局部窗口依此类推所以Deepseek基本上切断了这条依赖链只把最后128个tokens重放给每个解码7层这正是他们所说的SWA有借重放发挥作用的地方所以不是让100万个tokens经过全部40层而是让100万个经过前20层只让128个经过剩下的20层这意味着现在完整的预填" } ], "id": 373, "review_annotation": "", "row_id": 373, "bounded_row_id": 373, "parent_row_ids": [ 373 ], "parent_row_id": 373, "packet_id": 28, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 771.6635961538462, "packet_end": 800.8515817307693, "raw_text": "第39层本身又依赖更早的局部窗口", "selected_text": "第39层本身又依赖更早的局部窗口", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 373 ], "packet_id": 28, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 778.8635961538463, "end": 779.4435961538462, "text": "依此类推", "raw_asr_text": "依此类推", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 28, "words": [ { "word": "依", "start": 778.8635961538463, "end": 779.0235961538463, "probability": 0.91064453125 }, { "word": "此", "start": 779.0235961538463, "end": 779.1435961538463, "probability": 0.99609375 }, { "word": "类", "start": 779.1435961538463, "end": 779.2635961538463, "probability": 0.99560546875 }, { "word": "推", "start": 779.2635961538463, "end": 779.4435961538462, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.13071968680933901, "compression_ratio": 1.3051948051948052, "no_speech_prob": 2.576132709930512e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "不断累积第40层的SWA状态依赖第39层第39层本身又依赖更早的局部窗口依此类推所以Deepseek基本上切断了这条依赖链只把最后128个tokens重放给每个解码7层这正是他们所说的SWA有借重放发挥作用的地方所以不是让100万个tokens经过全部40层而是让100万个经过前20层只让128个经过剩下的20层这意味着现在完整的预填" } ], "id": 374, "review_annotation": "", "row_id": 374, "bounded_row_id": 374, "parent_row_ids": [ 374 ], "parent_row_id": 374, "packet_id": 28, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 771.6635961538462, "packet_end": 800.8515817307693, "raw_text": "依此类推", "selected_text": "依此类推", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 374 ], "packet_id": 28, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 779.4435961538462, "end": 782.6235961538463, "text": "所以Deepseek基本上切断了这条依赖链", "raw_asr_text": "所以Deepseek基本上切断了这条依赖链", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 28, "words": [ { "word": "所以", "start": 779.4435961538462, "end": 780.0235961538463, "probability": 0.9736328125 }, { "word": "De", "start": 780.0235961538463, "end": 780.2035961538462, "probability": 0.8076171875 }, { "word": "ep", "start": 780.2035961538462, "end": 780.3035961538462, "probability": 0.96875 }, { "word": "se", "start": 780.3035961538462, "end": 780.4435961538462, "probability": 0.3359375 }, { "word": "ek", "start": 780.4435961538462, "end": 780.5435961538462, "probability": 0.84228515625 }, { "word": "基本", "start": 780.5435961538462, "end": 780.8835961538463, "probability": 0.97802734375 }, { "word": "上", "start": 780.8835961538463, "end": 781.0835961538462, "probability": 0.97412109375 }, { "word": "切", "start": 781.0835961538462, "end": 781.3835961538463, "probability": 0.998046875 }, { "word": "断", "start": 781.3835961538463, "end": 781.6835961538462, "probability": 0.916015625 }, { "word": "了", "start": 781.6835961538462, "end": 781.8035961538462, "probability": 0.9931640625 }, { "word": "这", "start": 781.8035961538462, "end": 781.9635961538462, "probability": 0.9677734375 }, { "word": "条", "start": 781.9635961538462, "end": 782.1435961538463, "probability": 0.818359375 }, { "word": "依", "start": 782.1435961538463, "end": 782.3235961538462, "probability": 0.998046875 }, { "word": "赖", "start": 782.3235961538462, "end": 782.4435961538462, "probability": 0.9990234375 }, { "word": "链", "start": 782.4435961538462, "end": 782.6235961538463, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.13071968680933901, "compression_ratio": 1.3051948051948052, "no_speech_prob": 2.576132709930512e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "不断累积第40层的SWA状态依赖第39层第39层本身又依赖更早的局部窗口依此类推所以Deepseek基本上切断了这条依赖链只把最后128个tokens重放给每个解码7层这正是他们所说的SWA有借重放发挥作用的地方所以不是让100万个tokens经过全部40层而是让100万个经过前20层只让128个经过剩下的20层这意味着现在完整的预填" } ], "id": 375, "review_annotation": "", "row_id": 375, "bounded_row_id": 375, "parent_row_ids": [ 375 ], "parent_row_id": 375, "packet_id": 28, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 771.6635961538462, "packet_end": 800.8515817307693, "raw_text": "所以Deepseek基本上切断了这条依赖链", "selected_text": "所以Deepseek基本上切断了这条依赖链", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 375 ], "packet_id": 28, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 782.6235961538463, "end": 787.2635961538463, "text": "只把最后128个tokens重放给每个解码7层", "raw_asr_text": "只把最后128个tokens重放给每个解码7层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 28, "words": [ { "word": "只", "start": 782.6235961538463, "end": 782.8635961538463, "probability": 0.9794921875 }, { "word": "把", "start": 782.8635961538463, "end": 783.0635961538462, "probability": 0.9970703125 }, { "word": "最", "start": 783.0635961538462, "end": 783.3035961538462, "probability": 0.9853515625 }, { "word": "后", "start": 783.3035961538462, "end": 783.5635961538462, "probability": 0.99755859375 }, { "word": "12", "start": 783.5635961538462, "end": 784.0035961538463, "probability": 0.98974609375 }, { "word": "8", "start": 784.0035961538463, "end": 784.9235961538462, "probability": 0.99853515625 }, { "word": "个", "start": 784.9235961538462, "end": 785.0635961538462, "probability": 0.9853515625 }, { "word": "t", "start": 785.0635961538462, "end": 785.2235961538462, "probability": 0.59521484375 }, { "word": "ok", "start": 785.2235961538462, "end": 785.3435961538462, "probability": 0.6220703125 }, { "word": "ens", "start": 785.3435961538462, "end": 785.5435961538462, "probability": 0.98876953125 }, { "word": "重", "start": 785.5435961538462, "end": 785.8235961538462, "probability": 0.96142578125 }, { "word": "放", "start": 785.8235961538462, "end": 786.0235961538463, "probability": 0.978515625 }, { "word": "给", "start": 786.0235961538463, "end": 786.1835961538462, "probability": 0.9921875 }, { "word": "每", "start": 786.1835961538462, "end": 786.4035961538463, "probability": 0.9912109375 }, { "word": "个", "start": 786.4035961538463, "end": 786.5835961538462, "probability": 0.99755859375 }, { "word": "解", "start": 786.5835961538462, "end": 786.7635961538463, "probability": 0.96826171875 }, { "word": "码", "start": 786.7635961538463, "end": 786.9235961538462, "probability": 0.998046875 }, { "word": "7", "start": 786.9235961538462, "end": 787.0835961538462, "probability": 0.7470703125 }, { "word": "层", "start": 787.0835961538462, "end": 787.2635961538463, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.13071968680933901, "compression_ratio": 1.3051948051948052, "no_speech_prob": 2.576132709930512e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "不断累积第40层的SWA状态依赖第39层第39层本身又依赖更早的局部窗口依此类推所以Deepseek基本上切断了这条依赖链只把最后128个tokens重放给每个解码7层这正是他们所说的SWA有借重放发挥作用的地方所以不是让100万个tokens经过全部40层而是让100万个经过前20层只让128个经过剩下的20层这意味着现在完整的预填" } ], "id": 376, "review_annotation": "", "row_id": 376, "bounded_row_id": 376, "parent_row_ids": [ 376 ], "parent_row_id": 376, "packet_id": 28, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 771.6635961538462, "packet_end": 800.8515817307693, "raw_text": "只把最后128个tokens重放给每个解码7层", "selected_text": "只把最后128个tokens重放给每个解码7层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 376 ], "packet_id": 28, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 787.2635961538463, "end": 788.5835961538462, "text": "这正是他们所说的", "raw_asr_text": "这正是他们所说的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 28, "words": [ { "word": "这", "start": 787.2635961538463, "end": 787.4435961538462, "probability": 0.9755859375 }, { "word": "正", "start": 787.4435961538462, "end": 787.6035961538463, "probability": 0.99462890625 }, { "word": "是", "start": 787.6035961538463, "end": 787.7635961538463, "probability": 0.998046875 }, { "word": "他们", "start": 787.7635961538463, "end": 787.9635961538462, "probability": 0.9716796875 }, { "word": "所", "start": 787.9635961538462, "end": 788.2035961538462, "probability": 0.99755859375 }, { "word": "说", "start": 788.2035961538462, "end": 788.4035961538463, "probability": 0.9990234375 }, { "word": "的", "start": 788.4035961538463, "end": 788.5835961538462, "probability": 0.99609375 } ], "asr_metrics": { "avg_logprob": -0.13071968680933901, "compression_ratio": 1.3051948051948052, "no_speech_prob": 2.576132709930512e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "不断累积第40层的SWA状态依赖第39层第39层本身又依赖更早的局部窗口依此类推所以Deepseek基本上切断了这条依赖链只把最后128个tokens重放给每个解码7层这正是他们所说的SWA有借重放发挥作用的地方所以不是让100万个tokens经过全部40层而是让100万个经过前20层只让128个经过剩下的20层这意味着现在完整的预填" } ], "id": 377, "review_annotation": "", "row_id": 377, "bounded_row_id": 377, "parent_row_ids": [ 377 ], "parent_row_id": 377, "packet_id": 28, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 771.6635961538462, "packet_end": 800.8515817307693, "raw_text": "这正是他们所说的", "selected_text": "这正是他们所说的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 377 ], "packet_id": 28, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 788.5835961538462, "end": 791.2435961538463, "text": "SWA有借重放发挥作用的地方", "raw_asr_text": "SWA有借重放发挥作用的地方", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 28, "words": [ { "word": "S", "start": 788.5835961538462, "end": 788.8835961538463, "probability": 0.93505859375 }, { "word": "WA", "start": 788.8835961538463, "end": 789.3435961538462, "probability": 0.99462890625 }, { "word": "有", "start": 789.3435961538462, "end": 789.6235961538463, "probability": 0.958984375 }, { "word": "借", "start": 789.6235961538463, "end": 789.8235961538462, "probability": 0.81298828125 }, { "word": "重", "start": 789.8235961538462, "end": 790.0035961538463, "probability": 0.99560546875 }, { "word": "放", "start": 790.0035961538463, "end": 790.2235961538462, "probability": 0.99658203125 }, { "word": "发", "start": 790.2235961538462, "end": 790.3835961538463, "probability": 0.9912109375 }, { "word": "挥", "start": 790.3835961538463, "end": 790.5635961538462, "probability": 0.9990234375 }, { "word": "作", "start": 790.5635961538462, "end": 790.7235961538462, "probability": 0.98974609375 }, { "word": "用", "start": 790.7235961538462, "end": 790.8635961538463, "probability": 1.0 }, { "word": "的", "start": 790.8635961538463, "end": 790.9635961538462, "probability": 0.99853515625 }, { "word": "地方", "start": 790.9635961538462, "end": 791.2435961538463, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.13071968680933901, "compression_ratio": 1.3051948051948052, "no_speech_prob": 2.576132709930512e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "不断累积第40层的SWA状态依赖第39层第39层本身又依赖更早的局部窗口依此类推所以Deepseek基本上切断了这条依赖链只把最后128个tokens重放给每个解码7层这正是他们所说的SWA有借重放发挥作用的地方所以不是让100万个tokens经过全部40层而是让100万个经过前20层只让128个经过剩下的20层这意味着现在完整的预填" } ], "id": 378, "review_annotation": "", "row_id": 378, "bounded_row_id": 378, "parent_row_ids": [ 378 ], "parent_row_id": 378, "packet_id": 28, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 771.6635961538462, "packet_end": 800.8515817307693, "raw_text": "SWA有借重放发挥作用的地方", "selected_text": "SWA有借重放发挥作用的地方", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 378 ], "packet_id": 28, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 791.2435961538463, "end": 794.7435961538463, "text": "所以不是让100万个tokens经过全部40层", "raw_asr_text": "所以不是让100万个tokens经过全部40层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 28, "words": [ { "word": "所以", "start": 791.2435961538463, "end": 791.5435961538462, "probability": 0.98974609375 }, { "word": "不是", "start": 791.5435961538462, "end": 791.8635961538463, "probability": 0.98291015625 }, { "word": "让", "start": 791.8635961538463, "end": 792.0835961538462, "probability": 0.99169921875 }, { "word": "100", "start": 792.0835961538462, "end": 792.3635961538463, "probability": 0.9677734375 }, { "word": "万", "start": 792.3635961538463, "end": 792.6035961538463, "probability": 0.98046875 }, { "word": "个", "start": 792.6035961538463, "end": 792.8235961538462, "probability": 0.99169921875 }, { "word": "t", "start": 792.8235961538462, "end": 793.0435961538462, "probability": 0.95654296875 }, { "word": "ok", "start": 793.0435961538462, "end": 793.1435961538463, "probability": 0.97509765625 }, { "word": "ens", "start": 793.1435961538463, "end": 793.3435961538462, "probability": 0.998046875 }, { "word": "经", "start": 793.3435961538462, "end": 793.5635961538462, "probability": 0.96142578125 }, { "word": "过", "start": 793.5635961538462, "end": 793.8235961538462, "probability": 0.9990234375 }, { "word": "全部", "start": 793.8235961538462, "end": 794.1635961538462, "probability": 0.97265625 }, { "word": "40", "start": 794.1635961538462, "end": 794.5035961538463, "probability": 0.990234375 }, { "word": "层", "start": 794.5035961538463, "end": 794.7435961538463, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.13071968680933901, "compression_ratio": 1.3051948051948052, "no_speech_prob": 2.576132709930512e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "不断累积第40层的SWA状态依赖第39层第39层本身又依赖更早的局部窗口依此类推所以Deepseek基本上切断了这条依赖链只把最后128个tokens重放给每个解码7层这正是他们所说的SWA有借重放发挥作用的地方所以不是让100万个tokens经过全部40层而是让100万个经过前20层只让128个经过剩下的20层这意味着现在完整的预填" } ], "id": 379, "review_annotation": "", "row_id": 379, "bounded_row_id": 379, "parent_row_ids": [ 379 ], "parent_row_id": 379, "packet_id": 28, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 771.6635961538462, "packet_end": 800.8515817307693, "raw_text": "所以不是让100万个tokens经过全部40层", "selected_text": "所以不是让100万个tokens经过全部40层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 379 ], "packet_id": 28, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 794.7435961538463, "end": 796.8835961538463, "text": "而是让100万个经过前20层", "raw_asr_text": "而是让100万个经过前20层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 28, "words": [ { "word": "而", "start": 794.7435961538463, "end": 794.9035961538463, "probability": 0.9873046875 }, { "word": "是", "start": 794.9035961538463, "end": 795.0435961538462, "probability": 0.9990234375 }, { "word": "让", "start": 795.0435961538462, "end": 795.1835961538462, "probability": 0.99560546875 }, { "word": "100", "start": 795.1835961538462, "end": 795.4235961538462, "probability": 0.99365234375 }, { "word": "万", "start": 795.4235961538462, "end": 795.6235961538463, "probability": 0.9951171875 }, { "word": "个", "start": 795.6235961538463, "end": 795.7835961538462, "probability": 0.95361328125 }, { "word": "经", "start": 795.7835961538462, "end": 795.9235961538462, "probability": 0.939453125 }, { "word": "过", "start": 795.9235961538462, "end": 796.0835961538462, "probability": 0.99951171875 }, { "word": "前", "start": 796.0835961538462, "end": 796.3435961538462, "probability": 0.99658203125 }, { "word": "20", "start": 796.3435961538462, "end": 796.6235961538463, "probability": 0.98486328125 }, { "word": "层", "start": 796.6235961538463, "end": 796.8835961538463, "probability": 0.99609375 } ], "asr_metrics": { "avg_logprob": -0.13071968680933901, "compression_ratio": 1.3051948051948052, "no_speech_prob": 2.576132709930512e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "不断累积第40层的SWA状态依赖第39层第39层本身又依赖更早的局部窗口依此类推所以Deepseek基本上切断了这条依赖链只把最后128个tokens重放给每个解码7层这正是他们所说的SWA有借重放发挥作用的地方所以不是让100万个tokens经过全部40层而是让100万个经过前20层只让128个经过剩下的20层这意味着现在完整的预填" } ], "id": 380, "review_annotation": "", "row_id": 380, "bounded_row_id": 380, "parent_row_ids": [ 380 ], "parent_row_id": 380, "packet_id": 28, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 771.6635961538462, "packet_end": 800.8515817307693, "raw_text": "而是让100万个经过前20层", "selected_text": "而是让100万个经过前20层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 380 ], "packet_id": 28, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 796.8835961538463, "end": 799.1435961538463, "text": "只让128个经过剩下的20层", "raw_asr_text": "只让128个经过剩下的20层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 28, "words": [ { "word": "只", "start": 796.8835961538463, "end": 797.1035961538463, "probability": 0.96728515625 }, { "word": "让", "start": 797.1035961538463, "end": 797.2635961538463, "probability": 0.99755859375 }, { "word": "12", "start": 797.2635961538463, "end": 797.6435961538463, "probability": 0.98583984375 }, { "word": "8", "start": 797.6435961538463, "end": 797.9235961538462, "probability": 0.99951171875 }, { "word": "个", "start": 797.9235961538462, "end": 798.0835961538462, "probability": 0.98046875 }, { "word": "经", "start": 798.0835961538462, "end": 798.2635961538463, "probability": 0.92333984375 }, { "word": "过", "start": 798.2635961538463, "end": 798.3835961538463, "probability": 0.99755859375 }, { "word": "剩", "start": 798.3835961538463, "end": 798.5035961538463, "probability": 0.9951171875 }, { "word": "下", "start": 798.5035961538463, "end": 798.6835961538462, "probability": 0.99609375 }, { "word": "的", "start": 798.6835961538462, "end": 798.7435961538463, "probability": 0.98193359375 }, { "word": "20", "start": 798.7435961538463, "end": 798.9235961538462, "probability": 0.9951171875 }, { "word": "层", "start": 798.9235961538462, "end": 799.1435961538463, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.13071968680933901, "compression_ratio": 1.3051948051948052, "no_speech_prob": 2.576132709930512e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "不断累积第40层的SWA状态依赖第39层第39层本身又依赖更早的局部窗口依此类推所以Deepseek基本上切断了这条依赖链只把最后128个tokens重放给每个解码7层这正是他们所说的SWA有借重放发挥作用的地方所以不是让100万个tokens经过全部40层而是让100万个经过前20层只让128个经过剩下的20层这意味着现在完整的预填" } ], "id": 381, "review_annotation": "", "row_id": 381, "bounded_row_id": 381, "parent_row_ids": [ 381 ], "parent_row_id": 381, "packet_id": 28, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 771.6635961538462, "packet_end": 800.8515817307693, "raw_text": "只让128个经过剩下的20层", "selected_text": "只让128个经过剩下的20层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 381 ], "packet_id": 28, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 799.1435961538463, "end": 800.0515817307694, "text": "这意味着现在完整的预填", "raw_asr_text": "这意味着现在完整的预填", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 28, "words": [ { "word": "这", "start": 799.1435961538463, "end": 799.3635961538463, "probability": 0.98486328125 }, { "word": "意", "start": 799.3635961538463, "end": 799.4835961538463, "probability": 0.99462890625 }, { "word": "味", "start": 799.4835961538463, "end": 799.5635961538462, "probability": 0.99658203125 }, { "word": "着", "start": 799.5635961538462, "end": 799.6635961538462, "probability": 0.99560546875 }, { "word": "现在", "start": 799.6635961538462, "end": 799.8835961538463, "probability": 0.9921875 }, { "word": "完", "start": 799.8835961538463, "end": 800.1835961538462, "probability": 0.994140625 }, { "word": "整", "start": 800.1835961538462, "end": 800.3835961538463, "probability": 0.99951171875 }, { "word": "的", "start": 800.3835961538463, "end": 800.4835961538463, "probability": 0.97998046875 }, { "word": "预", "start": 800.4835961538463, "end": 800.6435961538463, "probability": 0.73974609375 }, { "word": "填", "start": 800.6435961538463, "end": 800.8235961538462, "probability": 0.77783203125 } ], "asr_metrics": { "avg_logprob": -0.13071968680933901, "compression_ratio": 1.3051948051948052, "no_speech_prob": 2.576132709930512e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "不断累积第40层的SWA状态依赖第39层第39层本身又依赖更早的局部窗口依此类推所以Deepseek基本上切断了这条依赖链只把最后128个tokens重放给每个解码7层这正是他们所说的SWA有借重放发挥作用的地方所以不是让100万个tokens经过全部40层而是让100万个经过前20层只让128个经过剩下的20层这意味着现在完整的预填" } ], "id": 382, "review_annotation": "", "row_id": 382, "bounded_row_id": 382, "parent_row_ids": [ 382 ], "parent_row_id": 382, "packet_id": 28, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 771.6635961538462, "packet_end": 800.8515817307693, "raw_text": "这意味着现在完整的预填", "selected_text": "这意味着现在完整的预填", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 382 ], "packet_id": 28, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 800.0515817307694, "end": 801.2915817307694, "text": "这意味着现在完整的预填充成本", "raw_asr_text": "这意味着现在完整的预填充成本", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 29, "words": [ { "word": "这", "start": 799.2515817307694, "end": 799.3715817307694, "probability": 0.97607421875 }, { "word": "意", "start": 799.3715817307694, "end": 799.5115817307694, "probability": 0.99072265625 }, { "word": "味", "start": 799.5115817307694, "end": 799.5715817307695, "probability": 0.99609375 }, { "word": "着", "start": 799.5715817307695, "end": 799.6715817307694, "probability": 0.9912109375 }, { "word": "现在", "start": 799.6715817307694, "end": 799.8715817307694, "probability": 0.9736328125 }, { "word": "完", "start": 799.8715817307694, "end": 800.1915817307695, "probability": 0.98486328125 }, { "word": "整", "start": 800.1915817307695, "end": 800.3715817307694, "probability": 0.99951171875 }, { "word": "的", "start": 800.3715817307694, "end": 800.4715817307695, "probability": 0.98876953125 }, { "word": "预", "start": 800.4715817307695, "end": 800.6515817307694, "probability": 0.97265625 }, { "word": "填", "start": 800.6515817307694, "end": 800.8315817307695, "probability": 0.94921875 }, { "word": "充", "start": 800.8315817307695, "end": 800.9715817307695, "probability": 0.9970703125 }, { "word": "成", "start": 800.9715817307695, "end": 801.1315817307694, "probability": 0.9990234375 }, { "word": "本", "start": 801.1315817307694, "end": 801.2915817307694, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09916735132899854, "compression_ratio": 1.335195530726257, "no_speech_prob": 1.9815994331540132e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这意味着现在完整的预填充成本大致等于完整提示词经过一半模型再加上128个tokens经过其余各层在100万tokens上下文的情况下那128个tokens的重放基本可以忽略不计所以整体预填充仍接近减半虽说只运行这128个tokens重建出的状态严格来说是近似的但Deepseek仍表示对质量的影响微乎其微而且有借重放实际上还带来了另一个相当大的好处因为Deepseek基本意识到已经不值得" } ], "id": 383, "review_annotation": "", "row_id": 383, "bounded_row_id": 383, "parent_row_ids": [ 383 ], "parent_row_id": 383, "packet_id": 29, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 799.2515817307694, "packet_end": 828.4395673076923, "raw_text": "这意味着现在完整的预填充成本", "selected_text": "这意味着现在完整的预填充成本", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 383 ], "packet_id": 29, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 801.2915817307694, "end": 803.9715817307695, "text": "大致等于完整提示词经过一半模型", "raw_asr_text": "大致等于完整提示词经过一半模型", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 29, "words": [ { "word": "大", "start": 801.2915817307694, "end": 801.5915817307695, "probability": 0.94873046875 }, { "word": "致", "start": 801.5915817307695, "end": 801.7715817307694, "probability": 0.9990234375 }, { "word": "等", "start": 801.7715817307694, "end": 801.8915817307694, "probability": 0.99169921875 }, { "word": "于", "start": 801.8915817307694, "end": 802.0515817307694, "probability": 0.99755859375 }, { "word": "完", "start": 802.0515817307694, "end": 802.2315817307694, "probability": 0.97412109375 }, { "word": "整", "start": 802.2315817307694, "end": 802.4515817307695, "probability": 0.99951171875 }, { "word": "提", "start": 802.4515817307695, "end": 802.6315817307694, "probability": 0.998046875 }, { "word": "示", "start": 802.6315817307694, "end": 802.7715817307694, "probability": 0.9990234375 }, { "word": "词", "start": 802.7715817307694, "end": 802.9515817307695, "probability": 0.994140625 }, { "word": "经", "start": 802.9515817307695, "end": 803.0915817307695, "probability": 0.9697265625 }, { "word": "过", "start": 803.0915817307695, "end": 803.2515817307694, "probability": 0.99853515625 }, { "word": "一", "start": 803.2515817307694, "end": 803.4115817307694, "probability": 0.9921875 }, { "word": "半", "start": 803.4115817307694, "end": 803.5515817307694, "probability": 0.98779296875 }, { "word": "模", "start": 803.5515817307694, "end": 803.7515817307694, "probability": 0.9521484375 }, { "word": "型", "start": 803.7515817307694, "end": 803.9715817307695, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.09916735132899854, "compression_ratio": 1.335195530726257, "no_speech_prob": 1.9815994331540132e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这意味着现在完整的预填充成本大致等于完整提示词经过一半模型再加上128个tokens经过其余各层在100万tokens上下文的情况下那128个tokens的重放基本可以忽略不计所以整体预填充仍接近减半虽说只运行这128个tokens重建出的状态严格来说是近似的但Deepseek仍表示对质量的影响微乎其微而且有借重放实际上还带来了另一个相当大的好处因为Deepseek基本意识到已经不值得" } ], "id": 384, "review_annotation": "", "row_id": 384, "bounded_row_id": 384, "parent_row_ids": [ 384 ], "parent_row_id": 384, "packet_id": 29, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 799.2515817307694, "packet_end": 828.4395673076923, "raw_text": "大致等于完整提示词经过一半模型", "selected_text": "大致等于完整提示词经过一半模型", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 384 ], "packet_id": 29, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 803.9715817307695, "end": 805.7515817307694, "text": "再加上128个tokens", "raw_asr_text": "再加上128个tokens", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 29, "words": [ { "word": "再", "start": 803.9715817307695, "end": 804.2115817307695, "probability": 0.90625 }, { "word": "加", "start": 804.2115817307695, "end": 804.3715817307694, "probability": 0.9990234375 }, { "word": "上", "start": 804.3715817307694, "end": 804.5115817307694, "probability": 0.97900390625 }, { "word": "12", "start": 804.5115817307694, "end": 804.9915817307694, "probability": 0.9833984375 }, { "word": "8", "start": 804.9915817307694, "end": 805.2715817307694, "probability": 0.99951171875 }, { "word": "个", "start": 805.2715817307694, "end": 805.3315817307695, "probability": 0.97802734375 }, { "word": "t", "start": 805.3315817307695, "end": 805.4515817307695, "probability": 0.64306640625 }, { "word": "ok", "start": 805.4515817307695, "end": 805.5515817307694, "probability": 0.6962890625 }, { "word": "ens", "start": 805.5515817307694, "end": 805.7515817307694, "probability": 0.98486328125 } ], "asr_metrics": { "avg_logprob": -0.09916735132899854, "compression_ratio": 1.335195530726257, "no_speech_prob": 1.9815994331540132e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这意味着现在完整的预填充成本大致等于完整提示词经过一半模型再加上128个tokens经过其余各层在100万tokens上下文的情况下那128个tokens的重放基本可以忽略不计所以整体预填充仍接近减半虽说只运行这128个tokens重建出的状态严格来说是近似的但Deepseek仍表示对质量的影响微乎其微而且有借重放实际上还带来了另一个相当大的好处因为Deepseek基本意识到已经不值得" } ], "id": 385, "review_annotation": "", "row_id": 385, "bounded_row_id": 385, "parent_row_ids": [ 385 ], "parent_row_id": 385, "packet_id": 29, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 799.2515817307694, "packet_end": 828.4395673076923, "raw_text": "再加上128个tokens", "selected_text": "再加上128个tokens", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 385 ], "packet_id": 29, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 805.7515817307694, "end": 807.0915817307695, "text": "经过其余各层", "raw_asr_text": "经过其余各层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 29, "words": [ { "word": "经", "start": 805.7515817307694, "end": 806.1115817307694, "probability": 0.9033203125 }, { "word": "过", "start": 806.1115817307694, "end": 806.2315817307694, "probability": 0.99853515625 }, { "word": "其", "start": 806.2315817307694, "end": 806.4715817307695, "probability": 0.99658203125 }, { "word": "余", "start": 806.4715817307695, "end": 806.7315817307694, "probability": 0.998046875 }, { "word": "各", "start": 806.7315817307694, "end": 806.8915817307694, "probability": 0.955078125 }, { "word": "层", "start": 806.8915817307694, "end": 807.0915817307695, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09916735132899854, "compression_ratio": 1.335195530726257, "no_speech_prob": 1.9815994331540132e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这意味着现在完整的预填充成本大致等于完整提示词经过一半模型再加上128个tokens经过其余各层在100万tokens上下文的情况下那128个tokens的重放基本可以忽略不计所以整体预填充仍接近减半虽说只运行这128个tokens重建出的状态严格来说是近似的但Deepseek仍表示对质量的影响微乎其微而且有借重放实际上还带来了另一个相当大的好处因为Deepseek基本意识到已经不值得" } ], "id": 386, "review_annotation": "", "row_id": 386, "bounded_row_id": 386, "parent_row_ids": [ 386 ], "parent_row_id": 386, "packet_id": 29, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 799.2515817307694, "packet_end": 828.4395673076923, "raw_text": "经过其余各层", "selected_text": "经过其余各层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 386 ], "packet_id": 29, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 807.0915817307695, "end": 809.5715817307695, "text": "在100万tokens上下文的情况下", "raw_asr_text": "在100万tokens上下文的情况下", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 29, "words": [ { "word": "在", "start": 807.0915817307695, "end": 807.2715817307694, "probability": 0.96533203125 }, { "word": "100", "start": 807.2715817307694, "end": 807.5515817307694, "probability": 0.96875 }, { "word": "万", "start": 807.5515817307694, "end": 807.7715817307694, "probability": 0.970703125 }, { "word": "t", "start": 807.7715817307694, "end": 807.9715817307695, "probability": 0.95947265625 }, { "word": "ok", "start": 807.9715817307695, "end": 808.1115817307694, "probability": 0.64013671875 }, { "word": "ens", "start": 808.1115817307694, "end": 808.2515817307694, "probability": 0.994140625 }, { "word": "上", "start": 808.2515817307694, "end": 808.4915817307694, "probability": 0.98828125 }, { "word": "下", "start": 808.4915817307694, "end": 808.6915817307695, "probability": 0.99853515625 }, { "word": "文", "start": 808.6915817307695, "end": 808.9115817307694, "probability": 0.6748046875 }, { "word": "的", "start": 808.9115817307694, "end": 809.0515817307694, "probability": 0.99267578125 }, { "word": "情况", "start": 809.0515817307694, "end": 809.2915817307694, "probability": 0.998046875 }, { "word": "下", "start": 809.2915817307694, "end": 809.5715817307695, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.09916735132899854, "compression_ratio": 1.335195530726257, "no_speech_prob": 1.9815994331540132e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这意味着现在完整的预填充成本大致等于完整提示词经过一半模型再加上128个tokens经过其余各层在100万tokens上下文的情况下那128个tokens的重放基本可以忽略不计所以整体预填充仍接近减半虽说只运行这128个tokens重建出的状态严格来说是近似的但Deepseek仍表示对质量的影响微乎其微而且有借重放实际上还带来了另一个相当大的好处因为Deepseek基本意识到已经不值得" } ], "id": 387, "review_annotation": "", "row_id": 387, "bounded_row_id": 387, "parent_row_ids": [ 387 ], "parent_row_id": 387, "packet_id": 29, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 799.2515817307694, "packet_end": 828.4395673076923, "raw_text": "在100万tokens上下文的情况下", "selected_text": "在100万tokens上下文的情况下", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 387 ], "packet_id": 29, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 809.5715817307695, "end": 811.2915817307694, "text": "那128个tokens的重放", "raw_asr_text": "那128个tokens的重放", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 29, "words": [ { "word": "那", "start": 809.5715817307695, "end": 809.7515817307694, "probability": 0.2410888671875 }, { "word": "12", "start": 809.7515817307694, "end": 810.1115817307694, "probability": 0.98974609375 }, { "word": "8", "start": 810.1115817307694, "end": 810.4315817307694, "probability": 0.99951171875 }, { "word": "个", "start": 810.4315817307694, "end": 810.5515817307694, "probability": 0.99169921875 }, { "word": "t", "start": 810.5515817307694, "end": 810.6715817307694, "probability": 0.98583984375 }, { "word": "ok", "start": 810.6715817307694, "end": 810.7515817307694, "probability": 0.8447265625 }, { "word": "ens", "start": 810.7515817307694, "end": 810.8715817307694, "probability": 0.99169921875 }, { "word": "的", "start": 810.8715817307694, "end": 810.9515817307695, "probability": 0.966796875 }, { "word": "重", "start": 810.9515817307695, "end": 811.0915817307695, "probability": 0.94189453125 }, { "word": "放", "start": 811.0915817307695, "end": 811.2915817307694, "probability": 0.966796875 } ], "asr_metrics": { "avg_logprob": -0.09916735132899854, "compression_ratio": 1.335195530726257, "no_speech_prob": 1.9815994331540132e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这意味着现在完整的预填充成本大致等于完整提示词经过一半模型再加上128个tokens经过其余各层在100万tokens上下文的情况下那128个tokens的重放基本可以忽略不计所以整体预填充仍接近减半虽说只运行这128个tokens重建出的状态严格来说是近似的但Deepseek仍表示对质量的影响微乎其微而且有借重放实际上还带来了另一个相当大的好处因为Deepseek基本意识到已经不值得" } ], "id": 388, "review_annotation": "", "row_id": 388, "bounded_row_id": 388, "parent_row_ids": [ 388 ], "parent_row_id": 388, "packet_id": 29, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 799.2515817307694, "packet_end": 828.4395673076923, "raw_text": "那128个tokens的重放", "selected_text": "那128个tokens的重放", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 388 ], "packet_id": 29, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 811.2915817307694, "end": 812.3715817307694, "text": "基本可以忽略不计", "raw_asr_text": "基本可以忽略不计", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 29, "words": [ { "word": "基本", "start": 811.2915817307694, "end": 811.5315817307694, "probability": 0.9892578125 }, { "word": "可以", "start": 811.5315817307694, "end": 811.7915817307694, "probability": 0.98583984375 }, { "word": "忽", "start": 811.7915817307694, "end": 811.9715817307695, "probability": 0.9951171875 }, { "word": "略", "start": 811.9715817307695, "end": 812.0715817307695, "probability": 0.9990234375 }, { "word": "不", "start": 812.0715817307695, "end": 812.2115817307695, "probability": 0.9990234375 }, { "word": "计", "start": 812.2115817307695, "end": 812.3715817307694, "probability": 0.97021484375 } ], "asr_metrics": { "avg_logprob": -0.09916735132899854, "compression_ratio": 1.335195530726257, "no_speech_prob": 1.9815994331540132e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这意味着现在完整的预填充成本大致等于完整提示词经过一半模型再加上128个tokens经过其余各层在100万tokens上下文的情况下那128个tokens的重放基本可以忽略不计所以整体预填充仍接近减半虽说只运行这128个tokens重建出的状态严格来说是近似的但Deepseek仍表示对质量的影响微乎其微而且有借重放实际上还带来了另一个相当大的好处因为Deepseek基本意识到已经不值得" } ], "id": 389, "review_annotation": "", "row_id": 389, "bounded_row_id": 389, "parent_row_ids": [ 389 ], "parent_row_id": 389, "packet_id": 29, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 799.2515817307694, "packet_end": 828.4395673076923, "raw_text": "基本可以忽略不计", "selected_text": "基本可以忽略不计", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 389 ], "packet_id": 29, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 812.3715817307694, "end": 814.9715817307695, "text": "所以整体预填充仍接近减半", "raw_asr_text": "所以整体预填充仍接近减半", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 29, "words": [ { "word": "所以", "start": 812.3715817307694, "end": 812.7515817307694, "probability": 0.98095703125 }, { "word": "整", "start": 812.7515817307694, "end": 812.9715817307695, "probability": 0.99609375 }, { "word": "体", "start": 812.9715817307695, "end": 813.1715817307694, "probability": 0.9990234375 }, { "word": "预", "start": 813.1715817307694, "end": 813.2915817307694, "probability": 0.9921875 }, { "word": "填", "start": 813.2915817307694, "end": 813.4515817307695, "probability": 0.9990234375 }, { "word": "充", "start": 813.4515817307695, "end": 813.6715817307694, "probability": 1.0 }, { "word": "仍", "start": 813.6715817307694, "end": 814.0115817307694, "probability": 0.99658203125 }, { "word": "接", "start": 814.0115817307694, "end": 814.2315817307694, "probability": 0.99755859375 }, { "word": "近", "start": 814.2315817307694, "end": 814.4515817307695, "probability": 0.99951171875 }, { "word": "减", "start": 814.4515817307695, "end": 814.7315817307694, "probability": 0.99609375 }, { "word": "半", "start": 814.7315817307694, "end": 814.9715817307695, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.09916735132899854, "compression_ratio": 1.335195530726257, "no_speech_prob": 1.9815994331540132e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这意味着现在完整的预填充成本大致等于完整提示词经过一半模型再加上128个tokens经过其余各层在100万tokens上下文的情况下那128个tokens的重放基本可以忽略不计所以整体预填充仍接近减半虽说只运行这128个tokens重建出的状态严格来说是近似的但Deepseek仍表示对质量的影响微乎其微而且有借重放实际上还带来了另一个相当大的好处因为Deepseek基本意识到已经不值得" } ], "id": 390, "review_annotation": "", "row_id": 390, "bounded_row_id": 390, "parent_row_ids": [ 390 ], "parent_row_id": 390, "packet_id": 29, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 799.2515817307694, "packet_end": 828.4395673076923, "raw_text": "所以整体预填充仍接近减半", "selected_text": "所以整体预填充仍接近减半", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 390 ], "packet_id": 29, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 814.9715817307695, "end": 817.2915817307694, "text": "虽说只运行这128个tokens", "raw_asr_text": "虽说只运行这128个tokens", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 29, "words": [ { "word": "虽", "start": 814.9715817307695, "end": 815.2515817307694, "probability": 0.93896484375 }, { "word": "说", "start": 815.2515817307694, "end": 815.3915817307694, "probability": 0.9814453125 }, { "word": "只", "start": 815.3915817307694, "end": 815.5715817307695, "probability": 0.98876953125 }, { "word": "运", "start": 815.5715817307695, "end": 815.7315817307694, "probability": 0.998046875 }, { "word": "行", "start": 815.7315817307694, "end": 815.8715817307694, "probability": 0.99853515625 }, { "word": "这", "start": 815.8715817307694, "end": 816.0515817307694, "probability": 0.9765625 }, { "word": "12", "start": 816.0515817307694, "end": 816.3115817307694, "probability": 0.9970703125 }, { "word": "8", "start": 816.3115817307694, "end": 816.6315817307694, "probability": 0.99951171875 }, { "word": "个", "start": 816.6315817307694, "end": 816.7515817307694, "probability": 0.99267578125 }, { "word": "t", "start": 816.7515817307694, "end": 816.8915817307694, "probability": 0.99560546875 }, { "word": "ok", "start": 816.8915817307694, "end": 817.0115817307694, "probability": 0.9951171875 }, { "word": "ens", "start": 817.0115817307694, "end": 817.2915817307694, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09916735132899854, "compression_ratio": 1.335195530726257, "no_speech_prob": 1.9815994331540132e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这意味着现在完整的预填充成本大致等于完整提示词经过一半模型再加上128个tokens经过其余各层在100万tokens上下文的情况下那128个tokens的重放基本可以忽略不计所以整体预填充仍接近减半虽说只运行这128个tokens重建出的状态严格来说是近似的但Deepseek仍表示对质量的影响微乎其微而且有借重放实际上还带来了另一个相当大的好处因为Deepseek基本意识到已经不值得" } ], "id": 391, "review_annotation": "", "row_id": 391, "bounded_row_id": 391, "parent_row_ids": [ 391 ], "parent_row_id": 391, "packet_id": 29, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 799.2515817307694, "packet_end": 828.4395673076923, "raw_text": "虽说只运行这128个tokens", "selected_text": "虽说只运行这128个tokens", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 391 ], "packet_id": 29, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 817.2915817307694, "end": 819.6315817307694, "text": "重建出的状态严格来说是近似的", "raw_asr_text": "重建出的状态严格来说是近似的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 29, "words": [ { "word": "重", "start": 817.2915817307694, "end": 817.5715817307695, "probability": 0.98193359375 }, { "word": "建", "start": 817.5715817307695, "end": 817.7715817307694, "probability": 0.99365234375 }, { "word": "出", "start": 817.7715817307694, "end": 817.8915817307694, "probability": 0.978515625 }, { "word": "的", "start": 817.8915817307694, "end": 817.9915817307694, "probability": 0.99609375 }, { "word": "状", "start": 817.9915817307694, "end": 818.1515817307694, "probability": 0.998046875 }, { "word": "态", "start": 818.1515817307694, "end": 818.3315817307695, "probability": 1.0 }, { "word": "严", "start": 818.3315817307695, "end": 818.5115817307694, "probability": 0.9853515625 }, { "word": "格", "start": 818.5115817307694, "end": 818.6315817307694, "probability": 0.9990234375 }, { "word": "来", "start": 818.6315817307694, "end": 818.7515817307694, "probability": 0.9873046875 }, { "word": "说", "start": 818.7515817307694, "end": 818.8915817307694, "probability": 0.998046875 }, { "word": "是", "start": 818.8915817307694, "end": 819.0315817307694, "probability": 0.99072265625 }, { "word": "近", "start": 819.0315817307694, "end": 819.1915817307695, "probability": 0.99658203125 }, { "word": "似", "start": 819.1915817307695, "end": 819.4715817307695, "probability": 0.9990234375 }, { "word": "的", "start": 819.4715817307695, "end": 819.6315817307694, "probability": 0.99658203125 } ], "asr_metrics": { "avg_logprob": -0.09916735132899854, "compression_ratio": 1.335195530726257, "no_speech_prob": 1.9815994331540132e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这意味着现在完整的预填充成本大致等于完整提示词经过一半模型再加上128个tokens经过其余各层在100万tokens上下文的情况下那128个tokens的重放基本可以忽略不计所以整体预填充仍接近减半虽说只运行这128个tokens重建出的状态严格来说是近似的但Deepseek仍表示对质量的影响微乎其微而且有借重放实际上还带来了另一个相当大的好处因为Deepseek基本意识到已经不值得" } ], "id": 392, "review_annotation": "", "row_id": 392, "bounded_row_id": 392, "parent_row_ids": [ 392 ], "parent_row_id": 392, "packet_id": 29, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 799.2515817307694, "packet_end": 828.4395673076923, "raw_text": "重建出的状态严格来说是近似的", "selected_text": "重建出的状态严格来说是近似的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 392 ], "packet_id": 29, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 819.6315817307694, "end": 822.4315817307694, "text": "但Deepseek仍表示对质量的影响微乎其微", "raw_asr_text": "但Deepseek仍表示对质量的影响微乎其微", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 29, "words": [ { "word": "但", "start": 819.6315817307694, "end": 819.8115817307694, "probability": 0.986328125 }, { "word": "De", "start": 819.8115817307694, "end": 819.9315817307694, "probability": 0.68701171875 }, { "word": "ep", "start": 819.9315817307694, "end": 820.0315817307694, "probability": 0.962890625 }, { "word": "se", "start": 820.0315817307694, "end": 820.1515817307694, "probability": 0.369873046875 }, { "word": "ek", "start": 820.1515817307694, "end": 820.2315817307694, "probability": 0.8671875 }, { "word": "仍", "start": 820.2315817307694, "end": 820.4715817307695, "probability": 0.9853515625 }, { "word": "表示", "start": 820.4715817307695, "end": 820.7715817307694, "probability": 0.998046875 }, { "word": "对", "start": 820.7715817307694, "end": 820.9715817307695, "probability": 0.994140625 }, { "word": "质", "start": 820.9715817307695, "end": 821.1515817307694, "probability": 1.0 }, { "word": "量", "start": 821.1515817307694, "end": 821.2915817307694, "probability": 1.0 }, { "word": "的", "start": 821.2915817307694, "end": 821.3915817307694, "probability": 0.9697265625 }, { "word": "影", "start": 821.3915817307694, "end": 821.5715817307695, "probability": 0.9990234375 }, { "word": "响", "start": 821.5715817307695, "end": 821.7515817307694, "probability": 1.0 }, { "word": "微", "start": 821.7515817307694, "end": 821.9515817307695, "probability": 0.9921875 }, { "word": "乎", "start": 821.9515817307695, "end": 822.0915817307695, "probability": 0.99560546875 }, { "word": "其", "start": 822.0915817307695, "end": 822.2515817307694, "probability": 0.72900390625 }, { "word": "微", "start": 822.2515817307694, "end": 822.4315817307694, "probability": 0.9970703125 } ], "asr_metrics": { "avg_logprob": -0.09916735132899854, "compression_ratio": 1.335195530726257, "no_speech_prob": 1.9815994331540132e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这意味着现在完整的预填充成本大致等于完整提示词经过一半模型再加上128个tokens经过其余各层在100万tokens上下文的情况下那128个tokens的重放基本可以忽略不计所以整体预填充仍接近减半虽说只运行这128个tokens重建出的状态严格来说是近似的但Deepseek仍表示对质量的影响微乎其微而且有借重放实际上还带来了另一个相当大的好处因为Deepseek基本意识到已经不值得" } ], "id": 393, "review_annotation": "", "row_id": 393, "bounded_row_id": 393, "parent_row_ids": [ 393 ], "parent_row_id": 393, "packet_id": 29, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 799.2515817307694, "packet_end": 828.4395673076923, "raw_text": "但Deepseek仍表示对质量的影响微乎其微", "selected_text": "但Deepseek仍表示对质量的影响微乎其微", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 393 ], "packet_id": 29, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 822.4315817307694, "end": 824.4115817307694, "text": "而且有借重放实际上还带来了", "raw_asr_text": "而且有借重放实际上还带来了", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 29, "words": [ { "word": "而且", "start": 822.4315817307694, "end": 822.9315817307694, "probability": 0.99072265625 }, { "word": "有", "start": 822.9315817307694, "end": 823.1115817307694, "probability": 0.98583984375 }, { "word": "借", "start": 823.1115817307694, "end": 823.2915817307694, "probability": 0.9404296875 }, { "word": "重", "start": 823.2915817307694, "end": 823.4315817307694, "probability": 0.64697265625 }, { "word": "放", "start": 823.4315817307694, "end": 823.5915817307695, "probability": 0.98974609375 }, { "word": "实", "start": 823.5915817307695, "end": 823.7715817307694, "probability": 0.98486328125 }, { "word": "际", "start": 823.7715817307694, "end": 823.8715817307694, "probability": 0.9990234375 }, { "word": "上", "start": 823.8715817307694, "end": 823.9715817307695, "probability": 0.9892578125 }, { "word": "还", "start": 823.9715817307695, "end": 824.0715817307695, "probability": 0.99658203125 }, { "word": "带", "start": 824.0715817307695, "end": 824.2715817307694, "probability": 0.99951171875 }, { "word": "来了", "start": 824.2715817307694, "end": 824.4115817307694, "probability": 0.87158203125 } ], "asr_metrics": { "avg_logprob": -0.09916735132899854, "compression_ratio": 1.335195530726257, "no_speech_prob": 1.9815994331540132e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这意味着现在完整的预填充成本大致等于完整提示词经过一半模型再加上128个tokens经过其余各层在100万tokens上下文的情况下那128个tokens的重放基本可以忽略不计所以整体预填充仍接近减半虽说只运行这128个tokens重建出的状态严格来说是近似的但Deepseek仍表示对质量的影响微乎其微而且有借重放实际上还带来了另一个相当大的好处因为Deepseek基本意识到已经不值得" } ], "id": 394, "review_annotation": "", "row_id": 394, "bounded_row_id": 394, "parent_row_ids": [ 394 ], "parent_row_id": 394, "packet_id": 29, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 799.2515817307694, "packet_end": 828.4395673076923, "raw_text": "而且有借重放实际上还带来了", "selected_text": "而且有借重放实际上还带来了", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 394 ], "packet_id": 29, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 824.4115817307694, "end": 825.9715817307695, "text": "另一个相当大的好处", "raw_asr_text": "另一个相当大的好处", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 29, "words": [ { "word": "另", "start": 824.4115817307694, "end": 824.5915817307695, "probability": 0.9853515625 }, { "word": "一个", "start": 824.5915817307695, "end": 824.8115817307694, "probability": 0.99560546875 }, { "word": "相", "start": 824.8115817307694, "end": 825.0515817307694, "probability": 0.9970703125 }, { "word": "当", "start": 825.0515817307694, "end": 825.3115817307694, "probability": 0.9990234375 }, { "word": "大的", "start": 825.3115817307694, "end": 825.5515817307694, "probability": 0.98291015625 }, { "word": "好", "start": 825.5515817307694, "end": 825.7915817307694, "probability": 0.99951171875 }, { "word": "处", "start": 825.7915817307694, "end": 825.9715817307695, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09916735132899854, "compression_ratio": 1.335195530726257, "no_speech_prob": 1.9815994331540132e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这意味着现在完整的预填充成本大致等于完整提示词经过一半模型再加上128个tokens经过其余各层在100万tokens上下文的情况下那128个tokens的重放基本可以忽略不计所以整体预填充仍接近减半虽说只运行这128个tokens重建出的状态严格来说是近似的但Deepseek仍表示对质量的影响微乎其微而且有借重放实际上还带来了另一个相当大的好处因为Deepseek基本意识到已经不值得" } ], "id": 395, "review_annotation": "", "row_id": 395, "bounded_row_id": 395, "parent_row_ids": [ 395 ], "parent_row_id": 395, "packet_id": 29, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 799.2515817307694, "packet_end": 828.4395673076923, "raw_text": "另一个相当大的好处", "selected_text": "另一个相当大的好处", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 395 ], "packet_id": 29, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 825.9715817307695, "end": 827.6395673076923, "text": "因为Deepseek基本意识到", "raw_asr_text": "因为Deepseek基本意识到", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 29, "words": [ { "word": "因为", "start": 825.9715817307695, "end": 826.2715817307694, "probability": 0.9921875 }, { "word": "De", "start": 826.2715817307694, "end": 826.4715817307695, "probability": 0.99755859375 }, { "word": "ep", "start": 826.4715817307695, "end": 826.5915817307695, "probability": 0.99951171875 }, { "word": "se", "start": 826.5915817307695, "end": 826.7715817307694, "probability": 0.97998046875 }, { "word": "ek", "start": 826.7715817307694, "end": 826.9315817307694, "probability": 1.0 }, { "word": "基本", "start": 826.9315817307694, "end": 827.2915817307694, "probability": 0.9716796875 }, { "word": "意", "start": 827.2915817307694, "end": 827.4315817307694, "probability": 0.99755859375 }, { "word": "识", "start": 827.4315817307694, "end": 827.5315817307694, "probability": 0.99853515625 }, { "word": "到", "start": 827.5315817307694, "end": 827.6915817307695, "probability": 0.994140625 } ], "asr_metrics": { "avg_logprob": -0.09916735132899854, "compression_ratio": 1.335195530726257, "no_speech_prob": 1.9815994331540132e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这意味着现在完整的预填充成本大致等于完整提示词经过一半模型再加上128个tokens经过其余各层在100万tokens上下文的情况下那128个tokens的重放基本可以忽略不计所以整体预填充仍接近减半虽说只运行这128个tokens重建出的状态严格来说是近似的但Deepseek仍表示对质量的影响微乎其微而且有借重放实际上还带来了另一个相当大的好处因为Deepseek基本意识到已经不值得" } ], "id": 396, "review_annotation": "", "row_id": 396, "bounded_row_id": 396, "parent_row_ids": [ 396 ], "parent_row_id": 396, "packet_id": 29, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 799.2515817307694, "packet_end": 828.4395673076923, "raw_text": "因为Deepseek基本意识到", "selected_text": "因为Deepseek基本意识到", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 396 ], "packet_id": 29, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 827.6395673076923, "end": 829.9195673076924, "text": "基本意识到已经不值得再把SWA-KV缓存", "raw_asr_text": "基本意识到已经不值得再把SWA-KV缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 30, "words": [ { "word": "基本", "start": 826.8395673076924, "end": 827.2395673076924, "probability": 0.9501953125 }, { "word": "意", "start": 827.2395673076924, "end": 827.4395673076924, "probability": 0.9990234375 }, { "word": "识", "start": 827.4395673076924, "end": 827.5395673076924, "probability": 0.98388671875 }, { "word": "到", "start": 827.5395673076924, "end": 827.6795673076924, "probability": 0.99462890625 }, { "word": "已经", "start": 827.6795673076924, "end": 827.8995673076923, "probability": 0.943359375 }, { "word": "不", "start": 827.8995673076923, "end": 828.1395673076923, "probability": 0.9951171875 }, { "word": "值", "start": 828.1395673076923, "end": 828.2995673076924, "probability": 0.9990234375 }, { "word": "得", "start": 828.2995673076924, "end": 828.4395673076924, "probability": 0.99951171875 }, { "word": "再", "start": 828.4395673076924, "end": 828.5795673076924, "probability": 0.9443359375 }, { "word": "把", "start": 828.5795673076924, "end": 828.7595673076924, "probability": 0.9951171875 }, { "word": "S", "start": 828.7595673076924, "end": 828.9995673076924, "probability": 0.9609375 }, { "word": "WA", "start": 828.9995673076924, "end": 829.3795673076924, "probability": 0.8759765625 }, { "word": "-", "start": 829.3795673076924, "end": 829.4395673076924, "probability": 0.6142578125 }, { "word": "K", "start": 829.4395673076924, "end": 829.4795673076924, "probability": 0.97021484375 }, { "word": "V", "start": 829.4795673076924, "end": 829.5995673076924, "probability": 0.99755859375 }, { "word": "缓", "start": 829.5995673076924, "end": 829.7595673076924, "probability": 0.9775390625 }, { "word": "存", "start": 829.7595673076924, "end": 829.9195673076924, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07553164468427594, "compression_ratio": 1.2729970326409497, "no_speech_prob": 2.372613217005437e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "基本意识到已经不值得再把SWA-KV缓存当作长期记忆对待了在Deepseek V4中当它们缓存前缀已共复用时固态硬盘上的持久化缓存会同时存储全局KV和SWA-KV缓存令人惊讶的是尽管SWA只覆盖一个很小的局部窗口那些未压缩逐层生成的SWA状态仍几乎占了持久化KV存储的一半除此之外这两种缓存的生命周期完全不同如果同一个长前缀被复用全球" } ], "id": 397, "review_annotation": "", "row_id": 397, "bounded_row_id": 397, "parent_row_ids": [ 397 ], "parent_row_id": 397, "packet_id": 30, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 826.8395673076924, "packet_end": 856.0275528846154, "raw_text": "基本意识到已经不值得再把SWA-KV缓存", "selected_text": "基本意识到已经不值得再把SWA-KV缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 397 ], "packet_id": 30, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 829.9195673076924, "end": 831.1995673076924, "text": "当作长期记忆对待了", "raw_asr_text": "当作长期记忆对待了", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 30, "words": [ { "word": "当", "start": 829.9195673076924, "end": 830.0795673076924, "probability": 0.95166015625 }, { "word": "作", "start": 830.0795673076924, "end": 830.2195673076924, "probability": 0.64794921875 }, { "word": "长", "start": 830.2195673076924, "end": 830.4195673076924, "probability": 0.99072265625 }, { "word": "期", "start": 830.4195673076924, "end": 830.5995673076924, "probability": 0.99951171875 }, { "word": "记", "start": 830.5995673076924, "end": 830.7195673076924, "probability": 0.98486328125 }, { "word": "忆", "start": 830.7195673076924, "end": 830.8195673076924, "probability": 0.9990234375 }, { "word": "对", "start": 830.8195673076924, "end": 830.9195673076924, "probability": 0.99853515625 }, { "word": "待", "start": 830.9195673076924, "end": 831.0995673076924, "probability": 0.98828125 }, { "word": "了", "start": 831.0995673076924, "end": 831.1995673076924, "probability": 0.97314453125 } ], "asr_metrics": { "avg_logprob": -0.07553164468427594, "compression_ratio": 1.2729970326409497, "no_speech_prob": 2.372613217005437e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "基本意识到已经不值得再把SWA-KV缓存当作长期记忆对待了在Deepseek V4中当它们缓存前缀已共复用时固态硬盘上的持久化缓存会同时存储全局KV和SWA-KV缓存令人惊讶的是尽管SWA只覆盖一个很小的局部窗口那些未压缩逐层生成的SWA状态仍几乎占了持久化KV存储的一半除此之外这两种缓存的生命周期完全不同如果同一个长前缀被复用全球" } ], "id": 398, "review_annotation": "", "row_id": 398, "bounded_row_id": 398, "parent_row_ids": [ 398 ], "parent_row_id": 398, "packet_id": 30, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 826.8395673076924, "packet_end": 856.0275528846154, "raw_text": "当作长期记忆对待了", "selected_text": "当作长期记忆对待了", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 398 ], "packet_id": 30, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 831.1995673076924, "end": 832.4195673076924, "text": "在Deepseek V4中", "raw_asr_text": "在Deepseek V4中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 30, "words": [ { "word": "在", "start": 831.1995673076924, "end": 831.4995673076924, "probability": 0.95361328125 }, { "word": "De", "start": 831.4995673076924, "end": 831.6795673076924, "probability": 0.60107421875 }, { "word": "ep", "start": 831.6795673076924, "end": 831.7595673076924, "probability": 0.779296875 }, { "word": "se", "start": 831.7595673076924, "end": 831.8595673076924, "probability": 0.66015625 }, { "word": "ek", "start": 831.8595673076924, "end": 831.9395673076924, "probability": 0.6318359375 }, { "word": " V", "start": 831.9395673076924, "end": 832.0795673076924, "probability": 0.5302734375 }, { "word": "4", "start": 832.0795673076924, "end": 832.2395673076924, "probability": 0.9951171875 }, { "word": "中", "start": 832.2395673076924, "end": 832.4195673076924, "probability": 0.85009765625 } ], "asr_metrics": { "avg_logprob": -0.07553164468427594, "compression_ratio": 1.2729970326409497, "no_speech_prob": 2.372613217005437e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "基本意识到已经不值得再把SWA-KV缓存当作长期记忆对待了在Deepseek V4中当它们缓存前缀已共复用时固态硬盘上的持久化缓存会同时存储全局KV和SWA-KV缓存令人惊讶的是尽管SWA只覆盖一个很小的局部窗口那些未压缩逐层生成的SWA状态仍几乎占了持久化KV存储的一半除此之外这两种缓存的生命周期完全不同如果同一个长前缀被复用全球" } ], "id": 399, "review_annotation": "", "row_id": 399, "bounded_row_id": 399, "parent_row_ids": [ 399 ], "parent_row_id": 399, "packet_id": 30, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 826.8395673076924, "packet_end": 856.0275528846154, "raw_text": "在Deepseek V4中", "selected_text": "在Deepseek V4中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 399 ], "packet_id": 30, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 832.4195673076924, "end": 834.5995673076924, "text": "当它们缓存前缀已共复用时", "raw_asr_text": "当它们缓存前缀已共复用时", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 30, "words": [ { "word": "当", "start": 832.4195673076924, "end": 832.6195673076924, "probability": 0.873046875 }, { "word": "它", "start": 832.6195673076924, "end": 832.7795673076924, "probability": 0.572265625 }, { "word": "们", "start": 832.7795673076924, "end": 832.8795673076924, "probability": 0.9970703125 }, { "word": "缓", "start": 832.8795673076924, "end": 833.1195673076924, "probability": 0.99951171875 }, { "word": "存", "start": 833.1195673076924, "end": 833.2795673076924, "probability": 1.0 }, { "word": "前", "start": 833.2795673076924, "end": 833.4995673076924, "probability": 0.978515625 }, { "word": "缀", "start": 833.4995673076924, "end": 833.6795673076924, "probability": 0.953125 }, { "word": "已", "start": 833.6795673076924, "end": 833.9195673076924, "probability": 0.7685546875 }, { "word": "共", "start": 833.9195673076924, "end": 834.0595673076924, "probability": 0.96875 }, { "word": "复", "start": 834.0595673076924, "end": 834.2195673076924, "probability": 0.95263671875 }, { "word": "用", "start": 834.2195673076924, "end": 834.3595673076924, "probability": 0.99951171875 }, { "word": "时", "start": 834.3595673076924, "end": 834.5995673076924, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.07553164468427594, "compression_ratio": 1.2729970326409497, "no_speech_prob": 2.372613217005437e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "基本意识到已经不值得再把SWA-KV缓存当作长期记忆对待了在Deepseek V4中当它们缓存前缀已共复用时固态硬盘上的持久化缓存会同时存储全局KV和SWA-KV缓存令人惊讶的是尽管SWA只覆盖一个很小的局部窗口那些未压缩逐层生成的SWA状态仍几乎占了持久化KV存储的一半除此之外这两种缓存的生命周期完全不同如果同一个长前缀被复用全球" } ], "id": 400, "review_annotation": "", "row_id": 400, "bounded_row_id": 400, "parent_row_ids": [ 400 ], "parent_row_id": 400, "packet_id": 30, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 826.8395673076924, "packet_end": 856.0275528846154, "raw_text": "当它们缓存前缀已共复用时", "selected_text": "当它们缓存前缀已共复用时", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 400 ], "packet_id": 30, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 834.5995673076924, "end": 836.3195673076924, "text": "固态硬盘上的持久化缓存", "raw_asr_text": "固态硬盘上的持久化缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 30, "words": [ { "word": "固", "start": 834.5995673076924, "end": 834.8395673076924, "probability": 0.9912109375 }, { "word": "态", "start": 834.8395673076924, "end": 834.9595673076924, "probability": 0.99169921875 }, { "word": "硬", "start": 834.9595673076924, "end": 835.1195673076924, "probability": 0.994140625 }, { "word": "盘", "start": 835.1195673076924, "end": 835.2395673076924, "probability": 0.98974609375 }, { "word": "上", "start": 835.2395673076924, "end": 835.3595673076924, "probability": 0.99169921875 }, { "word": "的", "start": 835.3595673076924, "end": 835.4595673076924, "probability": 0.99609375 }, { "word": "持", "start": 835.4595673076924, "end": 835.6795673076924, "probability": 0.99853515625 }, { "word": "久", "start": 835.6795673076924, "end": 835.8595673076924, "probability": 0.99951171875 }, { "word": "化", "start": 835.8595673076924, "end": 835.9795673076924, "probability": 0.98046875 }, { "word": "缓", "start": 835.9795673076924, "end": 836.1595673076924, "probability": 0.9990234375 }, { "word": "存", "start": 836.1595673076924, "end": 836.3195673076924, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07553164468427594, "compression_ratio": 1.2729970326409497, "no_speech_prob": 2.372613217005437e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "基本意识到已经不值得再把SWA-KV缓存当作长期记忆对待了在Deepseek V4中当它们缓存前缀已共复用时固态硬盘上的持久化缓存会同时存储全局KV和SWA-KV缓存令人惊讶的是尽管SWA只覆盖一个很小的局部窗口那些未压缩逐层生成的SWA状态仍几乎占了持久化KV存储的一半除此之外这两种缓存的生命周期完全不同如果同一个长前缀被复用全球" } ], "id": 401, "review_annotation": "", "row_id": 401, "bounded_row_id": 401, "parent_row_ids": [ 401 ], "parent_row_id": 401, "packet_id": 30, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 826.8395673076924, "packet_end": 856.0275528846154, "raw_text": "固态硬盘上的持久化缓存", "selected_text": "固态硬盘上的持久化缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 401 ], "packet_id": 30, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 836.3195673076924, "end": 839.8995673076923, "text": "会同时存储全局KV和SWA-KV缓存", "raw_asr_text": "会同时存储全局KV和SWA-KV缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 30, "words": [ { "word": "会", "start": 836.3195673076924, "end": 836.5195673076923, "probability": 0.99462890625 }, { "word": "同", "start": 836.5195673076923, "end": 836.7595673076924, "probability": 0.9990234375 }, { "word": "时", "start": 836.7595673076924, "end": 836.9995673076924, "probability": 0.99853515625 }, { "word": "存", "start": 836.9995673076924, "end": 837.1795673076924, "probability": 0.99951171875 }, { "word": "储", "start": 837.1795673076924, "end": 837.4195673076924, "probability": 0.9990234375 }, { "word": "全", "start": 837.4195673076924, "end": 837.7395673076924, "probability": 0.98193359375 }, { "word": "局", "start": 837.7395673076924, "end": 837.9195673076924, "probability": 0.990234375 }, { "word": "K", "start": 837.9195673076924, "end": 838.1395673076923, "probability": 0.93115234375 }, { "word": "V", "start": 838.1395673076923, "end": 838.3395673076924, "probability": 0.9990234375 }, { "word": "和", "start": 838.3395673076924, "end": 838.5995673076924, "probability": 0.904296875 }, { "word": "S", "start": 838.5995673076924, "end": 838.8995673076923, "probability": 0.99072265625 }, { "word": "WA", "start": 838.8995673076923, "end": 839.2795673076924, "probability": 0.9970703125 }, { "word": "-", "start": 839.2795673076924, "end": 839.3795673076924, "probability": 0.9921875 }, { "word": "K", "start": 839.3795673076924, "end": 839.3995673076923, "probability": 0.99755859375 }, { "word": "V", "start": 839.3995673076923, "end": 839.5395673076924, "probability": 0.99951171875 }, { "word": "缓", "start": 839.5395673076924, "end": 839.6795673076924, "probability": 0.9931640625 }, { "word": "存", "start": 839.6795673076924, "end": 839.8995673076923, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07553164468427594, "compression_ratio": 1.2729970326409497, "no_speech_prob": 2.372613217005437e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "基本意识到已经不值得再把SWA-KV缓存当作长期记忆对待了在Deepseek V4中当它们缓存前缀已共复用时固态硬盘上的持久化缓存会同时存储全局KV和SWA-KV缓存令人惊讶的是尽管SWA只覆盖一个很小的局部窗口那些未压缩逐层生成的SWA状态仍几乎占了持久化KV存储的一半除此之外这两种缓存的生命周期完全不同如果同一个长前缀被复用全球" } ], "id": 402, "review_annotation": "", "row_id": 402, "bounded_row_id": 402, "parent_row_ids": [ 402 ], "parent_row_id": 402, "packet_id": 30, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 826.8395673076924, "packet_end": 856.0275528846154, "raw_text": "会同时存储全局KV和SWA-KV缓存", "selected_text": "会同时存储全局KV和SWA-KV缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 402 ], "packet_id": 30, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 839.8995673076923, "end": 840.9795673076924, "text": "令人惊讶的是", "raw_asr_text": "令人惊讶的是", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 30, "words": [ { "word": "令", "start": 839.8995673076923, "end": 840.1795673076924, "probability": 0.96826171875 }, { "word": "人", "start": 840.1795673076924, "end": 840.3595673076924, "probability": 1.0 }, { "word": "惊", "start": 840.3595673076924, "end": 840.5195673076923, "probability": 0.99853515625 }, { "word": "讶", "start": 840.5195673076923, "end": 840.6795673076924, "probability": 0.9970703125 }, { "word": "的是", "start": 840.6795673076924, "end": 840.9795673076924, "probability": 0.99462890625 } ], "asr_metrics": { "avg_logprob": -0.07553164468427594, "compression_ratio": 1.2729970326409497, "no_speech_prob": 2.372613217005437e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "基本意识到已经不值得再把SWA-KV缓存当作长期记忆对待了在Deepseek V4中当它们缓存前缀已共复用时固态硬盘上的持久化缓存会同时存储全局KV和SWA-KV缓存令人惊讶的是尽管SWA只覆盖一个很小的局部窗口那些未压缩逐层生成的SWA状态仍几乎占了持久化KV存储的一半除此之外这两种缓存的生命周期完全不同如果同一个长前缀被复用全球" } ], "id": 403, "review_annotation": "", "row_id": 403, "bounded_row_id": 403, "parent_row_ids": [ 403 ], "parent_row_id": 403, "packet_id": 30, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 826.8395673076924, "packet_end": 856.0275528846154, "raw_text": "令人惊讶的是", "selected_text": "令人惊讶的是", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 403 ], "packet_id": 30, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 840.9795673076924, "end": 844.4395673076924, "text": "尽管SWA只覆盖一个很小的局部窗口", "raw_asr_text": "尽管SWA只覆盖一个很小的局部窗口", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 30, "words": [ { "word": "尽", "start": 840.9795673076924, "end": 841.3195673076924, "probability": 0.9638671875 }, { "word": "管", "start": 841.3195673076924, "end": 841.5195673076923, "probability": 0.99755859375 }, { "word": "S", "start": 841.5195673076923, "end": 841.7995673076924, "probability": 0.990234375 }, { "word": "WA", "start": 841.7995673076924, "end": 842.2195673076924, "probability": 0.9990234375 }, { "word": "只", "start": 842.2195673076924, "end": 842.4995673076924, "probability": 0.99169921875 }, { "word": "覆", "start": 842.4995673076924, "end": 842.6795673076924, "probability": 0.99755859375 }, { "word": "盖", "start": 842.6795673076924, "end": 842.8195673076924, "probability": 0.99609375 }, { "word": "一个", "start": 842.8195673076924, "end": 843.0195673076923, "probability": 0.98974609375 }, { "word": "很", "start": 843.0195673076923, "end": 843.3195673076924, "probability": 0.99462890625 }, { "word": "小", "start": 843.3195673076924, "end": 843.5795673076924, "probability": 0.99951171875 }, { "word": "的", "start": 843.5795673076924, "end": 843.7195673076924, "probability": 0.99951171875 }, { "word": "局", "start": 843.7195673076924, "end": 843.8795673076924, "probability": 0.99951171875 }, { "word": "部", "start": 843.8795673076924, "end": 844.0395673076924, "probability": 0.99853515625 }, { "word": "窗", "start": 844.0395673076924, "end": 844.1995673076924, "probability": 0.9990234375 }, { "word": "口", "start": 844.1995673076924, "end": 844.4395673076924, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.07553164468427594, "compression_ratio": 1.2729970326409497, "no_speech_prob": 2.372613217005437e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "基本意识到已经不值得再把SWA-KV缓存当作长期记忆对待了在Deepseek V4中当它们缓存前缀已共复用时固态硬盘上的持久化缓存会同时存储全局KV和SWA-KV缓存令人惊讶的是尽管SWA只覆盖一个很小的局部窗口那些未压缩逐层生成的SWA状态仍几乎占了持久化KV存储的一半除此之外这两种缓存的生命周期完全不同如果同一个长前缀被复用全球" } ], "id": 404, "review_annotation": "", "row_id": 404, "bounded_row_id": 404, "parent_row_ids": [ 404 ], "parent_row_id": 404, "packet_id": 30, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 826.8395673076924, "packet_end": 856.0275528846154, "raw_text": "尽管SWA只覆盖一个很小的局部窗口", "selected_text": "尽管SWA只覆盖一个很小的局部窗口", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 404 ], "packet_id": 30, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 844.4395673076924, "end": 847.5195673076923, "text": "那些未压缩逐层生成的SWA状态", "raw_asr_text": "那些未压缩逐层生成的SWA状态", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 30, "words": [ { "word": "那", "start": 844.4395673076924, "end": 844.6795673076924, "probability": 0.99462890625 }, { "word": "些", "start": 844.6795673076924, "end": 844.8595673076924, "probability": 0.99951171875 }, { "word": "未", "start": 844.8595673076924, "end": 845.0395673076924, "probability": 0.89404296875 }, { "word": "压", "start": 845.0395673076924, "end": 845.2795673076924, "probability": 0.9990234375 }, { "word": "缩", "start": 845.2795673076924, "end": 845.5195673076923, "probability": 0.99658203125 }, { "word": "逐", "start": 845.5195673076923, "end": 845.7195673076924, "probability": 0.9609375 }, { "word": "层", "start": 845.7195673076924, "end": 845.9395673076924, "probability": 0.99609375 }, { "word": "生", "start": 845.9395673076924, "end": 846.1195673076924, "probability": 0.9453125 }, { "word": "成", "start": 846.1195673076924, "end": 846.3195673076924, "probability": 0.84375 }, { "word": "的", "start": 846.3195673076924, "end": 846.4595673076924, "probability": 0.9990234375 }, { "word": "S", "start": 846.4595673076924, "end": 846.6795673076924, "probability": 0.99609375 }, { "word": "WA", "start": 846.6795673076924, "end": 847.0795673076924, "probability": 0.99951171875 }, { "word": "状", "start": 847.0795673076924, "end": 847.2995673076924, "probability": 0.986328125 }, { "word": "态", "start": 847.2995673076924, "end": 847.5195673076923, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.07553164468427594, "compression_ratio": 1.2729970326409497, "no_speech_prob": 2.372613217005437e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "基本意识到已经不值得再把SWA-KV缓存当作长期记忆对待了在Deepseek V4中当它们缓存前缀已共复用时固态硬盘上的持久化缓存会同时存储全局KV和SWA-KV缓存令人惊讶的是尽管SWA只覆盖一个很小的局部窗口那些未压缩逐层生成的SWA状态仍几乎占了持久化KV存储的一半除此之外这两种缓存的生命周期完全不同如果同一个长前缀被复用全球" } ], "id": 405, "review_annotation": "", "row_id": 405, "bounded_row_id": 405, "parent_row_ids": [ 405 ], "parent_row_id": 405, "packet_id": 30, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 826.8395673076924, "packet_end": 856.0275528846154, "raw_text": "那些未压缩逐层生成的SWA状态", "selected_text": "那些未压缩逐层生成的SWA状态", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 405 ], "packet_id": 30, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 847.5195673076923, "end": 850.3795673076924, "text": "仍几乎占了持久化KV存储的一半", "raw_asr_text": "仍几乎占了持久化KV存储的一半", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 30, "words": [ { "word": "仍", "start": 847.5195673076923, "end": 847.8195673076924, "probability": 0.9599609375 }, { "word": "几", "start": 847.8195673076924, "end": 848.1195673076924, "probability": 0.9951171875 }, { "word": "乎", "start": 848.1195673076924, "end": 848.2795673076924, "probability": 1.0 }, { "word": "占", "start": 848.2795673076924, "end": 848.4195673076924, "probability": 0.9951171875 }, { "word": "了", "start": 848.4195673076924, "end": 848.5995673076924, "probability": 0.99560546875 }, { "word": "持", "start": 848.5995673076924, "end": 848.8595673076924, "probability": 0.99658203125 }, { "word": "久", "start": 848.8595673076924, "end": 849.0995673076924, "probability": 1.0 }, { "word": "化", "start": 849.0995673076924, "end": 849.2795673076924, "probability": 0.99462890625 }, { "word": "K", "start": 849.2795673076924, "end": 849.4595673076924, "probability": 0.9833984375 }, { "word": "V", "start": 849.4595673076924, "end": 849.5795673076924, "probability": 0.99951171875 }, { "word": "存", "start": 849.5795673076924, "end": 849.7795673076924, "probability": 0.96923828125 }, { "word": "储", "start": 849.7795673076924, "end": 849.9195673076924, "probability": 0.9990234375 }, { "word": "的", "start": 849.9195673076924, "end": 850.0595673076924, "probability": 0.9912109375 }, { "word": "一", "start": 850.0595673076924, "end": 850.2195673076924, "probability": 0.9990234375 }, { "word": "半", "start": 850.2195673076924, "end": 850.3795673076924, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.07553164468427594, "compression_ratio": 1.2729970326409497, "no_speech_prob": 2.372613217005437e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "基本意识到已经不值得再把SWA-KV缓存当作长期记忆对待了在Deepseek V4中当它们缓存前缀已共复用时固态硬盘上的持久化缓存会同时存储全局KV和SWA-KV缓存令人惊讶的是尽管SWA只覆盖一个很小的局部窗口那些未压缩逐层生成的SWA状态仍几乎占了持久化KV存储的一半除此之外这两种缓存的生命周期完全不同如果同一个长前缀被复用全球" } ], "id": 406, "review_annotation": "", "row_id": 406, "bounded_row_id": 406, "parent_row_ids": [ 406 ], "parent_row_id": 406, "packet_id": 30, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 826.8395673076924, "packet_end": 856.0275528846154, "raw_text": "仍几乎占了持久化KV存储的一半", "selected_text": "仍几乎占了持久化KV存储的一半", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 406 ], "packet_id": 30, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 850.3795673076924, "end": 851.1195673076924, "text": "除此之外", "raw_asr_text": "除此之外", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 30, "words": [ { "word": "除", "start": 850.3795673076924, "end": 850.7195673076924, "probability": 0.99365234375 }, { "word": "此", "start": 850.7195673076924, "end": 850.8795673076924, "probability": 1.0 }, { "word": "之", "start": 850.8795673076924, "end": 850.9595673076924, "probability": 0.99951171875 }, { "word": "外", "start": 850.9595673076924, "end": 851.1195673076924, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07553164468427594, "compression_ratio": 1.2729970326409497, "no_speech_prob": 2.372613217005437e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "基本意识到已经不值得再把SWA-KV缓存当作长期记忆对待了在Deepseek V4中当它们缓存前缀已共复用时固态硬盘上的持久化缓存会同时存储全局KV和SWA-KV缓存令人惊讶的是尽管SWA只覆盖一个很小的局部窗口那些未压缩逐层生成的SWA状态仍几乎占了持久化KV存储的一半除此之外这两种缓存的生命周期完全不同如果同一个长前缀被复用全球" } ], "id": 407, "review_annotation": "", "row_id": 407, "bounded_row_id": 407, "parent_row_ids": [ 407 ], "parent_row_id": 407, "packet_id": 30, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 826.8395673076924, "packet_end": 856.0275528846154, "raw_text": "除此之外", "selected_text": "除此之外", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 407 ], "packet_id": 30, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 851.1195673076924, "end": 853.1995673076924, "text": "这两种缓存的生命周期完全不同", "raw_asr_text": "这两种缓存的生命周期完全不同", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 30, "words": [ { "word": "这", "start": 851.1195673076924, "end": 851.3195673076924, "probability": 0.97412109375 }, { "word": "两", "start": 851.3195673076924, "end": 851.4595673076924, "probability": 0.994140625 }, { "word": "种", "start": 851.4595673076924, "end": 851.5995673076924, "probability": 0.998046875 }, { "word": "缓", "start": 851.5995673076924, "end": 851.7795673076924, "probability": 0.9990234375 }, { "word": "存", "start": 851.7795673076924, "end": 851.9395673076924, "probability": 1.0 }, { "word": "的", "start": 851.9395673076924, "end": 852.0395673076924, "probability": 0.99560546875 }, { "word": "生", "start": 852.0395673076924, "end": 852.1595673076924, "probability": 0.99658203125 }, { "word": "命", "start": 852.1595673076924, "end": 852.2795673076924, "probability": 0.99951171875 }, { "word": "周", "start": 852.2795673076924, "end": 852.3995673076923, "probability": 0.998046875 }, { "word": "期", "start": 852.3995673076923, "end": 852.5595673076924, "probability": 0.99951171875 }, { "word": "完全", "start": 852.5595673076924, "end": 852.8795673076924, "probability": 0.99658203125 }, { "word": "不同", "start": 852.8795673076924, "end": 853.1995673076924, "probability": 0.99169921875 } ], "asr_metrics": { "avg_logprob": -0.07553164468427594, "compression_ratio": 1.2729970326409497, "no_speech_prob": 2.372613217005437e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "基本意识到已经不值得再把SWA-KV缓存当作长期记忆对待了在Deepseek V4中当它们缓存前缀已共复用时固态硬盘上的持久化缓存会同时存储全局KV和SWA-KV缓存令人惊讶的是尽管SWA只覆盖一个很小的局部窗口那些未压缩逐层生成的SWA状态仍几乎占了持久化KV存储的一半除此之外这两种缓存的生命周期完全不同如果同一个长前缀被复用全球" } ], "id": 408, "review_annotation": "", "row_id": 408, "bounded_row_id": 408, "parent_row_ids": [ 408 ], "parent_row_id": 408, "packet_id": 30, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 826.8395673076924, "packet_end": 856.0275528846154, "raw_text": "这两种缓存的生命周期完全不同", "selected_text": "这两种缓存的生命周期完全不同", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 408 ], "packet_id": 30, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 853.1995673076924, "end": 855.2275528846154, "text": "如果同一个长前缀被复用", "raw_asr_text": "如果同一个长前缀被复用", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 30, "words": [ { "word": "如果", "start": 853.1995673076924, "end": 853.8595673076924, "probability": 0.990234375 }, { "word": "同", "start": 853.8595673076924, "end": 854.1595673076924, "probability": 0.99609375 }, { "word": "一个", "start": 854.1595673076924, "end": 854.4195673076924, "probability": 0.99560546875 }, { "word": "长", "start": 854.4195673076924, "end": 854.6795673076924, "probability": 0.96044921875 }, { "word": "前", "start": 854.6795673076924, "end": 854.8995673076923, "probability": 0.9833984375 }, { "word": "缀", "start": 854.8995673076923, "end": 855.0395673076924, "probability": 0.9775390625 }, { "word": "被", "start": 855.0395673076924, "end": 855.2595673076924, "probability": 0.98876953125 }, { "word": "复", "start": 855.2595673076924, "end": 855.3995673076923, "probability": 0.95703125 }, { "word": "用", "start": 855.3995673076923, "end": 855.5795673076924, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.07553164468427594, "compression_ratio": 1.2729970326409497, "no_speech_prob": 2.372613217005437e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "基本意识到已经不值得再把SWA-KV缓存当作长期记忆对待了在Deepseek V4中当它们缓存前缀已共复用时固态硬盘上的持久化缓存会同时存储全局KV和SWA-KV缓存令人惊讶的是尽管SWA只覆盖一个很小的局部窗口那些未压缩逐层生成的SWA状态仍几乎占了持久化KV存储的一半除此之外这两种缓存的生命周期完全不同如果同一个长前缀被复用全球" } ], "id": 409, "review_annotation": "", "row_id": 409, "bounded_row_id": 409, "parent_row_ids": [ 409 ], "parent_row_id": 409, "packet_id": 30, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 826.8395673076924, "packet_end": 856.0275528846154, "raw_text": "如果同一个长前缀被复用", "selected_text": "如果同一个长前缀被复用", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 409 ], "packet_id": 30, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 855.5875528846154, "end": 856.8475528846154, "text": "全局KV缓存", "raw_asr_text": "全局KV缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "全", "start": 855.5875528846154, "end": 855.8875528846155, "probability": 0.91552734375 }, { "word": "局", "start": 855.8875528846155, "end": 856.0675528846155, "probability": 0.9990234375 }, { "word": "K", "start": 856.0675528846155, "end": 856.2475528846155, "probability": 0.61279296875 }, { "word": "V", "start": 856.2475528846155, "end": 856.4075528846155, "probability": 0.9951171875 }, { "word": "缓", "start": 856.4075528846155, "end": 856.6475528846155, "probability": 0.998046875 }, { "word": "存", "start": 856.6475528846155, "end": 856.8475528846154, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 410, "review_annotation": "", "row_id": 410, "bounded_row_id": 410, "parent_row_ids": [ 410 ], "parent_row_id": 410, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "全局KV缓存", "selected_text": "全局KV缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 410 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 856.8475528846154, "end": 858.6675528846155, "text": "可能在数小时或数天后", "raw_asr_text": "可能在数小时或数天后", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "可能", "start": 856.8475528846154, "end": 857.1675528846155, "probability": 0.9931640625 }, { "word": "在", "start": 857.1675528846155, "end": 857.3475528846154, "probability": 0.9990234375 }, { "word": "数", "start": 857.3475528846154, "end": 857.5475528846155, "probability": 0.99951171875 }, { "word": "小", "start": 857.5475528846155, "end": 857.7275528846154, "probability": 0.99609375 }, { "word": "时", "start": 857.7275528846154, "end": 857.9075528846155, "probability": 0.99951171875 }, { "word": "或", "start": 857.9075528846155, "end": 858.1075528846154, "probability": 0.9921875 }, { "word": "数", "start": 858.1075528846154, "end": 858.2875528846155, "probability": 1.0 }, { "word": "天", "start": 858.2875528846155, "end": 858.4475528846154, "probability": 0.99853515625 }, { "word": "后", "start": 858.4475528846154, "end": 858.6675528846155, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 411, "review_annotation": "", "row_id": 411, "bounded_row_id": 411, "parent_row_ids": [ 411 ], "parent_row_id": 411, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "可能在数小时或数天后", "selected_text": "可能在数小时或数天后", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 411 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 858.6675528846155, "end": 859.4275528846155, "text": "仍然有用", "raw_asr_text": "仍然有用", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "仍", "start": 858.6675528846155, "end": 858.9475528846154, "probability": 0.9990234375 }, { "word": "然", "start": 858.9475528846154, "end": 859.0675528846155, "probability": 0.9990234375 }, { "word": "有", "start": 859.0675528846155, "end": 859.2475528846155, "probability": 1.0 }, { "word": "用", "start": 859.2475528846155, "end": 859.4275528846155, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 412, "review_annotation": "", "row_id": 412, "bounded_row_id": 412, "parent_row_ids": [ 412 ], "parent_row_id": 412, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "仍然有用", "selected_text": "仍然有用", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 412 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 859.4275528846155, "end": 861.2475528846155, "text": "但SWAKV缓存", "raw_asr_text": "但SWAKV缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "但", "start": 859.4275528846155, "end": 859.7475528846155, "probability": 0.9794921875 }, { "word": "S", "start": 859.7475528846155, "end": 860.0475528846155, "probability": 0.9794921875 }, { "word": "W", "start": 860.0475528846155, "end": 860.2475528846155, "probability": 0.55322265625 }, { "word": "AK", "start": 860.2475528846155, "end": 860.5875528846154, "probability": 0.93212890625 }, { "word": "V", "start": 860.5875528846154, "end": 860.8275528846154, "probability": 0.99755859375 }, { "word": "缓", "start": 860.8275528846154, "end": 861.0275528846155, "probability": 1.0 }, { "word": "存", "start": 861.0275528846155, "end": 861.2475528846155, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 413, "review_annotation": "", "row_id": 413, "bounded_row_id": 413, "parent_row_ids": [ 413 ], "parent_row_id": 413, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "但SWAKV缓存", "selected_text": "但SWAKV缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 413 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 861.2475528846155, "end": 862.8675528846155, "text": "其实仅对活跃绘画的", "raw_asr_text": "其实仅对活跃绘画的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "其实", "start": 861.2475528846155, "end": 861.5875528846154, "probability": 0.99755859375 }, { "word": "仅", "start": 861.5875528846154, "end": 861.9075528846155, "probability": 0.99658203125 }, { "word": "对", "start": 861.9075528846155, "end": 862.0475528846155, "probability": 0.9990234375 }, { "word": "活", "start": 862.0475528846155, "end": 862.2475528846155, "probability": 0.99951171875 }, { "word": "跃", "start": 862.2475528846155, "end": 862.4275528846155, "probability": 0.998046875 }, { "word": "绘", "start": 862.4275528846155, "end": 862.5675528846155, "probability": 0.9580078125 }, { "word": "画", "start": 862.5675528846155, "end": 862.7075528846154, "probability": 0.75390625 }, { "word": "的", "start": 862.7075528846154, "end": 862.8675528846155, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 414, "review_annotation": "", "row_id": 414, "bounded_row_id": 414, "parent_row_ids": [ 414 ], "parent_row_id": 414, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "其实仅对活跃绘画的", "selected_text": "其实仅对活跃绘画的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 414 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 862.8675528846155, "end": 864.3475528846154, "text": "当前局部状态有用", "raw_asr_text": "当前局部状态有用", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "当", "start": 862.8675528846155, "end": 863.0475528846155, "probability": 0.99951171875 }, { "word": "前", "start": 863.0475528846155, "end": 863.3075528846155, "probability": 0.99951171875 }, { "word": "局", "start": 863.3075528846155, "end": 863.5075528846155, "probability": 0.99951171875 }, { "word": "部", "start": 863.5075528846155, "end": 863.7075528846154, "probability": 0.99462890625 }, { "word": "状", "start": 863.7075528846154, "end": 863.8875528846155, "probability": 1.0 }, { "word": "态", "start": 863.8875528846155, "end": 864.0275528846155, "probability": 0.99951171875 }, { "word": "有", "start": 864.0275528846155, "end": 864.1875528846155, "probability": 0.9990234375 }, { "word": "用", "start": 864.1875528846155, "end": 864.3475528846154, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 415, "review_annotation": "", "row_id": 415, "bounded_row_id": 415, "parent_row_ids": [ 415 ], "parent_row_id": 415, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "当前局部状态有用", "selected_text": "当前局部状态有用", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 415 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 864.3475528846154, "end": 866.0275528846155, "text": "有效期通常只有几分钟", "raw_asr_text": "有效期通常只有几分钟", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "有", "start": 864.3475528846154, "end": 864.6475528846155, "probability": 0.99169921875 }, { "word": "效", "start": 864.6475528846155, "end": 864.7675528846155, "probability": 0.99951171875 }, { "word": "期", "start": 864.7675528846155, "end": 864.9275528846155, "probability": 0.99951171875 }, { "word": "通", "start": 864.9275528846155, "end": 865.1275528846155, "probability": 0.998046875 }, { "word": "常", "start": 865.1275528846155, "end": 865.3075528846155, "probability": 1.0 }, { "word": "只有", "start": 865.3075528846155, "end": 865.5475528846155, "probability": 0.9970703125 }, { "word": "几", "start": 865.5475528846155, "end": 865.7275528846154, "probability": 1.0 }, { "word": "分", "start": 865.7275528846154, "end": 865.8875528846155, "probability": 0.998046875 }, { "word": "钟", "start": 865.8875528846155, "end": 866.0275528846155, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 416, "review_annotation": "", "row_id": 416, "bounded_row_id": 416, "parent_row_ids": [ 416 ], "parent_row_id": 416, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "有效期通常只有几分钟", "selected_text": "有效期通常只有几分钟", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 416 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 866.0275528846155, "end": 867.3675528846155, "text": "因此在V4.1中", "raw_asr_text": "因此在V4.1中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "因", "start": 866.0275528846155, "end": 866.2075528846154, "probability": 0.99755859375 }, { "word": "此", "start": 866.2075528846154, "end": 866.3675528846155, "probability": 1.0 }, { "word": "在", "start": 866.3675528846155, "end": 866.5475528846155, "probability": 0.99951171875 }, { "word": "V", "start": 866.5475528846155, "end": 866.7275528846154, "probability": 0.953125 }, { "word": "4.", "start": 866.7275528846154, "end": 866.9075528846155, "probability": 0.99072265625 }, { "word": "1", "start": 866.9875528846154, "end": 867.1675528846155, "probability": 0.99755859375 }, { "word": "中", "start": 867.1675528846155, "end": 867.3675528846155, "probability": 0.9921875 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 417, "review_annotation": "", "row_id": 417, "bounded_row_id": 417, "parent_row_ids": [ 417 ], "parent_row_id": 417, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "因此在V4.1中", "selected_text": "因此在V4.1中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 417 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 867.3675528846155, "end": 870.1075528846154, "text": "Deepseek完全不再把SWAKV缓存", "raw_asr_text": "Deepseek完全不再把SWAKV缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "De", "start": 867.3675528846155, "end": 867.6075528846154, "probability": 0.90283203125 }, { "word": "ep", "start": 867.6075528846154, "end": 867.7075528846154, "probability": 0.9775390625 }, { "word": "se", "start": 867.7075528846154, "end": 867.8475528846154, "probability": 0.419921875 }, { "word": "ek", "start": 867.8475528846154, "end": 867.9675528846154, "probability": 0.763671875 }, { "word": "完全", "start": 867.9675528846154, "end": 868.3275528846154, "probability": 0.99951171875 }, { "word": "不", "start": 868.3275528846154, "end": 868.6275528846155, "probability": 0.99951171875 }, { "word": "再", "start": 868.6275528846155, "end": 868.7675528846155, "probability": 0.99853515625 }, { "word": "把", "start": 868.7675528846155, "end": 868.9275528846155, "probability": 0.9990234375 }, { "word": "S", "start": 868.9275528846155, "end": 869.1275528846155, "probability": 0.99755859375 }, { "word": "W", "start": 869.1275528846155, "end": 869.2875528846155, "probability": 0.99853515625 }, { "word": "AK", "start": 869.2875528846155, "end": 869.5675528846155, "probability": 0.99853515625 }, { "word": "V", "start": 869.5675528846155, "end": 869.7275528846154, "probability": 1.0 }, { "word": "缓", "start": 869.7275528846154, "end": 869.9075528846155, "probability": 1.0 }, { "word": "存", "start": 869.9075528846155, "end": 870.1075528846154, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 418, "review_annotation": "", "row_id": 418, "bounded_row_id": 418, "parent_row_ids": [ 418 ], "parent_row_id": 418, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "Deepseek完全不再把SWAKV缓存", "selected_text": "Deepseek完全不再把SWAKV缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 418 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 870.1075528846154, "end": 872.0475528846155, "text": "存入长生命周期的持久化缓存", "raw_asr_text": "存入长生命周期的持久化缓存", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "存", "start": 870.1075528846154, "end": 870.3675528846155, "probability": 0.99951171875 }, { "word": "入", "start": 870.3675528846155, "end": 870.4875528846154, "probability": 0.99951171875 }, { "word": "长", "start": 870.4875528846154, "end": 870.6675528846155, "probability": 0.94873046875 }, { "word": "生", "start": 870.6675528846155, "end": 870.8675528846155, "probability": 1.0 }, { "word": "命", "start": 870.8675528846155, "end": 870.9875528846154, "probability": 1.0 }, { "word": "周", "start": 870.9875528846154, "end": 871.1075528846154, "probability": 0.99853515625 }, { "word": "期", "start": 871.1075528846154, "end": 871.2475528846155, "probability": 1.0 }, { "word": "的", "start": 871.2475528846155, "end": 871.3675528846155, "probability": 0.998046875 }, { "word": "持", "start": 871.3675528846155, "end": 871.4875528846154, "probability": 0.9990234375 }, { "word": "久", "start": 871.4875528846154, "end": 871.6275528846155, "probability": 0.9921875 }, { "word": "化", "start": 871.6275528846155, "end": 871.7275528846154, "probability": 0.9931640625 }, { "word": "缓", "start": 871.7275528846154, "end": 871.8875528846155, "probability": 1.0 }, { "word": "存", "start": 871.8875528846155, "end": 872.0475528846155, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 419, "review_annotation": "", "row_id": 419, "bounded_row_id": 419, "parent_row_ids": [ 419 ], "parent_row_id": 419, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "存入长生命周期的持久化缓存", "selected_text": "存入长生命周期的持久化缓存", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 419 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 872.0475528846155, "end": 873.9675528846154, "text": "而是将其临时存放在", "raw_asr_text": "而是将其临时存放在", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "而", "start": 872.0475528846155, "end": 872.1675528846155, "probability": 0.9990234375 }, { "word": "是", "start": 872.1675528846155, "end": 872.3075528846155, "probability": 0.99951171875 }, { "word": "将", "start": 872.3075528846155, "end": 872.4875528846154, "probability": 1.0 }, { "word": "其", "start": 872.4875528846154, "end": 872.7275528846154, "probability": 1.0 }, { "word": "临", "start": 872.7275528846154, "end": 873.0675528846155, "probability": 0.98974609375 }, { "word": "时", "start": 873.0675528846155, "end": 873.3275528846154, "probability": 0.99853515625 }, { "word": "存", "start": 873.3275528846154, "end": 873.5675528846155, "probability": 0.99951171875 }, { "word": "放", "start": 873.5675528846155, "end": 873.7875528846155, "probability": 0.99853515625 }, { "word": "在", "start": 873.7875528846155, "end": 873.9675528846154, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 420, "review_annotation": "", "row_id": 420, "bounded_row_id": 420, "parent_row_ids": [ 420 ], "parent_row_id": 420, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "而是将其临时存放在", "selected_text": "而是将其临时存放在", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 420 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 873.9675528846154, "end": 875.9275528846155, "text": "小得多的主机内存池中", "raw_asr_text": "小得多的主机内存池中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "小", "start": 873.9675528846154, "end": 874.2875528846155, "probability": 0.9970703125 }, { "word": "得", "start": 874.2875528846155, "end": 874.4475528846154, "probability": 0.78515625 }, { "word": "多", "start": 874.4475528846154, "end": 874.5875528846154, "probability": 1.0 }, { "word": "的", "start": 874.5875528846154, "end": 874.7875528846155, "probability": 0.99951171875 }, { "word": "主", "start": 874.7875528846155, "end": 874.9875528846154, "probability": 0.99609375 }, { "word": "机", "start": 874.9875528846154, "end": 875.1675528846155, "probability": 0.8994140625 }, { "word": "内", "start": 875.1675528846155, "end": 875.3075528846155, "probability": 0.99755859375 }, { "word": "存", "start": 875.3075528846155, "end": 875.4875528846154, "probability": 0.99951171875 }, { "word": "池", "start": 875.4875528846154, "end": 875.7275528846154, "probability": 0.9921875 }, { "word": "中", "start": 875.7275528846154, "end": 875.9275528846155, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 421, "review_annotation": "", "row_id": 421, "bounded_row_id": 421, "parent_row_ids": [ 421 ], "parent_row_id": 421, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "小得多的主机内存池中", "selected_text": "小得多的主机内存池中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 421 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 875.9275528846155, "end": 877.4675528846154, "text": "一旦过期或被驱逐", "raw_asr_text": "一旦过期或被驱逐", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "一", "start": 875.9275528846155, "end": 876.1875528846155, "probability": 0.9990234375 }, { "word": "旦", "start": 876.1875528846155, "end": 876.3475528846154, "probability": 0.9990234375 }, { "word": "过", "start": 876.3475528846154, "end": 876.4875528846154, "probability": 0.99951171875 }, { "word": "期", "start": 876.4875528846154, "end": 876.6475528846155, "probability": 0.966796875 }, { "word": "或", "start": 876.6475528846155, "end": 876.8275528846154, "probability": 0.974609375 }, { "word": "被", "start": 876.8275528846154, "end": 877.0075528846155, "probability": 0.99951171875 }, { "word": "驱", "start": 877.0075528846155, "end": 877.2675528846155, "probability": 0.998046875 }, { "word": "逐", "start": 877.2675528846155, "end": 877.4675528846154, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 422, "review_annotation": "", "row_id": 422, "bounded_row_id": 422, "parent_row_ids": [ 422 ], "parent_row_id": 422, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "一旦过期或被驱逐", "selected_text": "一旦过期或被驱逐", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 422 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 877.4675528846154, "end": 879.6075528846154, "text": "就直接用有界重放重建", "raw_asr_text": "就直接用有界重放重建", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "就", "start": 877.4675528846154, "end": 877.7075528846154, "probability": 0.99755859375 }, { "word": "直接", "start": 877.7075528846154, "end": 878.0675528846155, "probability": 1.0 }, { "word": "用", "start": 878.0675528846155, "end": 878.3275528846154, "probability": 0.99951171875 }, { "word": "有", "start": 878.3275528846154, "end": 878.6275528846155, "probability": 0.96826171875 }, { "word": "界", "start": 878.6275528846155, "end": 878.8275528846154, "probability": 0.369873046875 }, { "word": "重", "start": 878.8275528846154, "end": 879.0275528846155, "probability": 0.53759765625 }, { "word": "放", "start": 879.0275528846155, "end": 879.2075528846154, "probability": 0.953125 }, { "word": "重", "start": 879.2075528846154, "end": 879.4275528846155, "probability": 0.99365234375 }, { "word": "建", "start": 879.4275528846155, "end": 879.6075528846154, "probability": 0.9970703125 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 423, "review_annotation": "", "row_id": 423, "bounded_row_id": 423, "parent_row_ids": [ 423 ], "parent_row_id": 423, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "就直接用有界重放重建", "selected_text": "就直接用有界重放重建", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 423 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 879.6075528846154, "end": 880.9675528846154, "text": "最后128个Token", "raw_asr_text": "最后128个Token", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "最", "start": 879.6075528846154, "end": 879.8275528846154, "probability": 0.9970703125 }, { "word": "后", "start": 879.8275528846154, "end": 879.9675528846154, "probability": 0.99951171875 }, { "word": "12", "start": 879.9675528846154, "end": 880.2075528846154, "probability": 0.9970703125 }, { "word": "8", "start": 880.2075528846154, "end": 880.5275528846155, "probability": 0.99951171875 }, { "word": "个", "start": 880.5275528846155, "end": 880.6475528846155, "probability": 0.99609375 }, { "word": "T", "start": 880.6475528846155, "end": 880.7675528846155, "probability": 0.6484375 }, { "word": "oken", "start": 880.7675528846155, "end": 880.9675528846154, "probability": 0.6416015625 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 424, "review_annotation": "", "row_id": 424, "bounded_row_id": 424, "parent_row_ids": [ 424 ], "parent_row_id": 424, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "最后128个Token", "selected_text": "最后128个Token", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 424 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 880.9675528846154, "end": 881.8275528846154, "text": "再来说Token", "raw_asr_text": "再来说Token", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "再", "start": 880.9675528846154, "end": 881.3275528846154, "probability": 0.9150390625 }, { "word": "来", "start": 881.3275528846154, "end": 881.4275528846155, "probability": 1.0 }, { "word": "说", "start": 881.4275528846155, "end": 881.5275528846155, "probability": 0.99951171875 }, { "word": "T", "start": 881.5275528846155, "end": 881.6475528846155, "probability": 0.8984375 }, { "word": "oken", "start": 881.6475528846155, "end": 881.8275528846154, "probability": 0.89990234375 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 425, "review_annotation": "", "row_id": 425, "bounded_row_id": 425, "parent_row_ids": [ 425 ], "parent_row_id": 425, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "再来说Token", "selected_text": "再来说Token", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 425 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 881.8275528846154, "end": 882.7275528846154, "text": "与此同时", "raw_asr_text": "与此同时", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 31, "words": [ { "word": "与", "start": 881.8275528846154, "end": 882.2675528846155, "probability": 0.9853515625 }, { "word": "此", "start": 882.2675528846155, "end": 882.3475528846154, "probability": 0.998046875 }, { "word": "同", "start": 882.3475528846154, "end": 882.5075528846155, "probability": 1.0 }, { "word": "时", "start": 882.5075528846155, "end": 882.7275528846154, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09708673596958031, "compression_ratio": 1.2462908011869436, "no_speech_prob": 2.678894779617469e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "长前坠被复用全局KV缓存可能在数小时或数天后仍然有用但SWAKV缓存其实仅对活跃绘画的当前局部状态有用有效期通常只有几分钟因此在V4.1中Deepseek完全不再把SWAKV缓存存入长生命周期的持久化缓存而是将其临时存放在小得多的主机内存池中一旦过期或被驱逐就直接用有界重放重建最后128个Token再来说Token与此同时全局KV" } ], "id": 426, "review_annotation": "", "row_id": 426, "bounded_row_id": 426, "parent_row_ids": [ 426 ], "parent_row_id": 426, "packet_id": 31, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 854.4275528846155, "packet_end": 883.6155384615384, "raw_text": "与此同时", "selected_text": "与此同时", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 426 ], "packet_id": 31, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 882.8155384615385, "end": 886.2355384615386, "text": "与此同时全局KV可以存储在固态硬盘这类持久化存储中", "raw_asr_text": "与此同时全局KV可以存储在固态硬盘这类持久化存储中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 32, "words": [ { "word": "与", "start": 882.0155384615385, "end": 882.2755384615385, "probability": 0.861328125 }, { "word": "此", "start": 882.2755384615385, "end": 882.3555384615386, "probability": 0.9697265625 }, { "word": "同", "start": 882.3555384615386, "end": 882.4955384615386, "probability": 0.99853515625 }, { "word": "时", "start": 882.4955384615386, "end": 882.7155384615386, "probability": 0.99755859375 }, { "word": "全", "start": 882.7155384615386, "end": 883.0755384615385, "probability": 0.74658203125 }, { "word": "局", "start": 883.0755384615385, "end": 883.2155384615386, "probability": 0.9931640625 }, { "word": "K", "start": 883.2155384615386, "end": 883.4355384615385, "probability": 0.7900390625 }, { "word": "V", "start": 883.4355384615385, "end": 883.5955384615386, "probability": 0.994140625 }, { "word": "可以", "start": 883.5955384615386, "end": 883.8555384615386, "probability": 0.96142578125 }, { "word": "存", "start": 883.8555384615386, "end": 883.9955384615386, "probability": 0.998046875 }, { "word": "储", "start": 883.9955384615386, "end": 884.1355384615385, "probability": 0.998046875 }, { "word": "在", "start": 884.1355384615385, "end": 884.3555384615386, "probability": 0.99609375 }, { "word": "固", "start": 884.3555384615386, "end": 884.6155384615386, "probability": 0.9921875 }, { "word": "态", "start": 884.6155384615386, "end": 884.7355384615386, "probability": 0.994140625 }, { "word": "硬", "start": 884.7355384615386, "end": 884.8755384615386, "probability": 0.99951171875 }, { "word": "盘", "start": 884.8755384615386, "end": 885.0555384615385, "probability": 0.97412109375 }, { "word": "这", "start": 885.0555384615385, "end": 885.2355384615386, "probability": 0.98095703125 }, { "word": "类", "start": 885.2355384615386, "end": 885.3755384615386, "probability": 0.9990234375 }, { "word": "持", "start": 885.3755384615386, "end": 885.4955384615386, "probability": 0.998046875 }, { "word": "久", "start": 885.4955384615386, "end": 885.6555384615385, "probability": 0.99951171875 }, { "word": "化", "start": 885.6555384615385, "end": 885.7555384615385, "probability": 0.98193359375 }, { "word": "存", "start": 885.7555384615385, "end": 885.9155384615385, "probability": 0.998046875 }, { "word": "储", "start": 885.9155384615385, "end": 886.0555384615385, "probability": 0.99951171875 }, { "word": "中", "start": 886.0555384615385, "end": 886.2355384615386, "probability": 0.99072265625 } ], "asr_metrics": { "avg_logprob": -0.10354346613730153, "compression_ratio": 1.3265993265993266, "no_speech_prob": 1.9754167052132843e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "与此同时全局KV可以存储在固态硬盘这类持久化存储中这部分的持久化KV缩减幅度非常可观CSAR搭配FP4已经能把全局KV压缩到V4时期容量的四分之一左右而将SWAKV移出持久化存储后缓存规模几乎直接再减一半回看V4的设计持久化KV的占用空间大概一半来自SWA另一半来自全局KVV4.1首先通过CSAR和FP4把这占一半" } ], "id": 427, "review_annotation": "", "row_id": 427, "bounded_row_id": 427, "parent_row_ids": [ 427 ], "parent_row_id": 427, "packet_id": 32, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 882.0155384615385, "packet_end": 911.2035240384616, "raw_text": "与此同时全局KV可以存储在固态硬盘这类持久化存储中", "selected_text": "与此同时全局KV可以存储在固态硬盘这类持久化存储中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 427 ], "packet_id": 32, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 886.2355384615386, "end": 888.9955384615386, "text": "这部分的持久化KV缩减幅度非常可观", "raw_asr_text": "这部分的持久化KV缩减幅度非常可观", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 32, "words": [ { "word": "这", "start": 886.2355384615386, "end": 886.5355384615385, "probability": 0.93896484375 }, { "word": "部分", "start": 886.5355384615385, "end": 886.7755384615385, "probability": 0.9658203125 }, { "word": "的", "start": 886.7755384615385, "end": 886.8955384615385, "probability": 0.9462890625 }, { "word": "持", "start": 886.8955384615385, "end": 887.0355384615385, "probability": 0.9990234375 }, { "word": "久", "start": 887.0355384615385, "end": 887.1555384615385, "probability": 0.99951171875 }, { "word": "化", "start": 887.1555384615385, "end": 887.2755384615385, "probability": 0.9892578125 }, { "word": "K", "start": 887.2755384615385, "end": 887.4155384615385, "probability": 0.97314453125 }, { "word": "V", "start": 887.4155384615385, "end": 887.5355384615385, "probability": 0.9990234375 }, { "word": "缩", "start": 887.5355384615385, "end": 887.7555384615385, "probability": 0.9541015625 }, { "word": "减", "start": 887.7555384615385, "end": 887.9155384615385, "probability": 0.998046875 }, { "word": "幅", "start": 887.9155384615385, "end": 888.0955384615386, "probability": 0.99560546875 }, { "word": "度", "start": 888.0955384615386, "end": 888.2355384615386, "probability": 0.9990234375 }, { "word": "非常", "start": 888.2355384615386, "end": 888.5955384615386, "probability": 0.98291015625 }, { "word": "可", "start": 888.5955384615386, "end": 888.7755384615385, "probability": 0.9951171875 }, { "word": "观", "start": 888.7755384615385, "end": 888.9955384615386, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.10354346613730153, "compression_ratio": 1.3265993265993266, "no_speech_prob": 1.9754167052132843e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "与此同时全局KV可以存储在固态硬盘这类持久化存储中这部分的持久化KV缩减幅度非常可观CSAR搭配FP4已经能把全局KV压缩到V4时期容量的四分之一左右而将SWAKV移出持久化存储后缓存规模几乎直接再减一半回看V4的设计持久化KV的占用空间大概一半来自SWA另一半来自全局KVV4.1首先通过CSAR和FP4把这占一半" } ], "id": 428, "review_annotation": "", "row_id": 428, "bounded_row_id": 428, "parent_row_ids": [ 428 ], "parent_row_id": 428, "packet_id": 32, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 882.0155384615385, "packet_end": 911.2035240384616, "raw_text": "这部分的持久化KV缩减幅度非常可观", "selected_text": "这部分的持久化KV缩减幅度非常可观", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 428 ], "packet_id": 32, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 888.9955384615386, "end": 895.1155384615386, "text": "CSAR搭配FP4已经能把全局KV压缩到V4时期容量的四分之一左右", "raw_asr_text": "CSAR搭配FP4已经能把全局KV压缩到V4时期容量的四分之一左右", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 32, "words": [ { "word": "CS", "start": 888.9955384615386, "end": 889.3555384615386, "probability": 0.71728515625 }, { "word": "AR", "start": 889.3555384615386, "end": 889.6955384615385, "probability": 0.9599609375 }, { "word": "搭", "start": 889.6955384615385, "end": 890.0155384615385, "probability": 0.986328125 }, { "word": "配", "start": 890.0155384615385, "end": 890.1955384615385, "probability": 0.9990234375 }, { "word": "FP", "start": 890.1955384615385, "end": 890.5155384615385, "probability": 0.90283203125 }, { "word": "4", "start": 890.5155384615385, "end": 890.7955384615385, "probability": 0.9970703125 }, { "word": "已经", "start": 890.7955384615385, "end": 891.0955384615386, "probability": 0.96240234375 }, { "word": "能", "start": 891.0955384615386, "end": 891.2955384615385, "probability": 0.984375 }, { "word": "把", "start": 891.2955384615385, "end": 891.4155384615385, "probability": 0.99658203125 }, { "word": "全", "start": 891.4155384615385, "end": 891.7355384615386, "probability": 0.998046875 }, { "word": "局", "start": 891.7355384615386, "end": 891.9355384615385, "probability": 1.0 }, { "word": "K", "start": 891.9355384615385, "end": 892.1155384615386, "probability": 0.9921875 }, { "word": "V", "start": 892.1155384615386, "end": 892.3355384615386, "probability": 0.99951171875 }, { "word": "压", "start": 892.3355384615386, "end": 892.5755384615385, "probability": 0.98828125 }, { "word": "缩", "start": 892.5755384615385, "end": 892.7355384615386, "probability": 0.998046875 }, { "word": "到", "start": 892.7355384615386, "end": 892.8755384615386, "probability": 0.9921875 }, { "word": "V", "start": 892.8755384615386, "end": 893.0955384615386, "probability": 0.9375 }, { "word": "4", "start": 893.0955384615386, "end": 893.2955384615385, "probability": 0.98583984375 }, { "word": "时", "start": 893.2955384615385, "end": 893.4955384615386, "probability": 0.99169921875 }, { "word": "期", "start": 893.4955384615386, "end": 893.6555384615385, "probability": 0.99609375 }, { "word": "容", "start": 893.6555384615385, "end": 893.8155384615385, "probability": 0.99658203125 }, { "word": "量", "start": 893.8155384615385, "end": 894.0155384615385, "probability": 1.0 }, { "word": "的", "start": 894.0155384615385, "end": 894.1355384615385, "probability": 0.9833984375 }, { "word": "四", "start": 894.1355384615385, "end": 894.2955384615385, "probability": 0.51171875 }, { "word": "分", "start": 894.2955384615385, "end": 894.4355384615385, "probability": 0.9990234375 }, { "word": "之", "start": 894.4355384615385, "end": 894.5755384615385, "probability": 0.99755859375 }, { "word": "一", "start": 894.5755384615385, "end": 894.7555384615385, "probability": 0.97900390625 }, { "word": "左右", "start": 894.7555384615385, "end": 895.1155384615386, "probability": 0.99169921875 } ], "asr_metrics": { "avg_logprob": -0.10354346613730153, "compression_ratio": 1.3265993265993266, "no_speech_prob": 1.9754167052132843e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "与此同时全局KV可以存储在固态硬盘这类持久化存储中这部分的持久化KV缩减幅度非常可观CSAR搭配FP4已经能把全局KV压缩到V4时期容量的四分之一左右而将SWAKV移出持久化存储后缓存规模几乎直接再减一半回看V4的设计持久化KV的占用空间大概一半来自SWA另一半来自全局KVV4.1首先通过CSAR和FP4把这占一半" } ], "id": 429, "review_annotation": "", "row_id": 429, "bounded_row_id": 429, "parent_row_ids": [ 429 ], "parent_row_id": 429, "packet_id": 32, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 882.0155384615385, "packet_end": 911.2035240384616, "raw_text": "CSAR搭配FP4已经能把全局KV压缩到V4时期容量的四分之一左右", "selected_text": "CSAR搭配FP4已经能把全局KV压缩到V4时期容量的四分之一左右", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 429 ], "packet_id": 32, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 895.1155384615386, "end": 898.1355384615385, "text": "而将SWAKV移出持久化存储后", "raw_asr_text": "而将SWAKV移出持久化存储后", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 32, "words": [ { "word": "而", "start": 895.1155384615386, "end": 895.3555384615386, "probability": 0.9794921875 }, { "word": "将", "start": 895.3555384615386, "end": 895.5355384615385, "probability": 0.9931640625 }, { "word": "S", "start": 895.5355384615385, "end": 895.7555384615385, "probability": 0.9775390625 }, { "word": "WA", "start": 895.7555384615385, "end": 896.1555384615385, "probability": 0.9501953125 }, { "word": "K", "start": 896.1555384615385, "end": 896.3755384615386, "probability": 0.498046875 }, { "word": "V", "start": 896.3755384615386, "end": 896.5555384615385, "probability": 0.99951171875 }, { "word": "移", "start": 896.5555384615385, "end": 896.7555384615385, "probability": 0.9990234375 }, { "word": "出", "start": 896.7555384615385, "end": 896.9155384615385, "probability": 0.96337890625 }, { "word": "持", "start": 896.9155384615385, "end": 897.1755384615385, "probability": 0.9951171875 }, { "word": "久", "start": 897.1755384615385, "end": 897.3755384615386, "probability": 1.0 }, { "word": "化", "start": 897.3755384615386, "end": 897.5355384615385, "probability": 0.99658203125 }, { "word": "存", "start": 897.5355384615385, "end": 897.7555384615385, "probability": 0.998046875 }, { "word": "储", "start": 897.7555384615385, "end": 897.9355384615385, "probability": 1.0 }, { "word": "后", "start": 897.9355384615385, "end": 898.1355384615385, "probability": 0.98974609375 } ], "asr_metrics": { "avg_logprob": -0.10354346613730153, "compression_ratio": 1.3265993265993266, "no_speech_prob": 1.9754167052132843e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "与此同时全局KV可以存储在固态硬盘这类持久化存储中这部分的持久化KV缩减幅度非常可观CSAR搭配FP4已经能把全局KV压缩到V4时期容量的四分之一左右而将SWAKV移出持久化存储后缓存规模几乎直接再减一半回看V4的设计持久化KV的占用空间大概一半来自SWA另一半来自全局KVV4.1首先通过CSAR和FP4把这占一半" } ], "id": 430, "review_annotation": "", "row_id": 430, "bounded_row_id": 430, "parent_row_ids": [ 430 ], "parent_row_id": 430, "packet_id": 32, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 882.0155384615385, "packet_end": 911.2035240384616, "raw_text": "而将SWAKV移出持久化存储后", "selected_text": "而将SWAKV移出持久化存储后", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 430 ], "packet_id": 32, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 898.1355384615385, "end": 900.5155384615385, "text": "缓存规模几乎直接再减一半", "raw_asr_text": "缓存规模几乎直接再减一半", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 32, "words": [ { "word": "缓", "start": 898.1355384615385, "end": 898.4755384615386, "probability": 0.9921875 }, { "word": "存", "start": 898.4755384615386, "end": 898.6555384615385, "probability": 0.99951171875 }, { "word": "规", "start": 898.6555384615385, "end": 898.8355384615386, "probability": 0.99951171875 }, { "word": "模", "start": 898.8355384615386, "end": 898.9955384615386, "probability": 0.99951171875 }, { "word": "几", "start": 898.9955384615386, "end": 899.2755384615385, "probability": 0.99658203125 }, { "word": "乎", "start": 899.2755384615385, "end": 899.4555384615386, "probability": 1.0 }, { "word": "直接", "start": 899.4555384615386, "end": 899.7755384615385, "probability": 0.9814453125 }, { "word": "再", "start": 899.7755384615385, "end": 900.0355384615385, "probability": 0.84228515625 }, { "word": "减", "start": 900.0355384615385, "end": 900.2355384615386, "probability": 0.986328125 }, { "word": "一", "start": 900.2355384615386, "end": 900.4155384615385, "probability": 0.99658203125 }, { "word": "半", "start": 900.4155384615385, "end": 900.5155384615385, "probability": 0.99560546875 } ], "asr_metrics": { "avg_logprob": -0.10354346613730153, "compression_ratio": 1.3265993265993266, "no_speech_prob": 1.9754167052132843e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "与此同时全局KV可以存储在固态硬盘这类持久化存储中这部分的持久化KV缩减幅度非常可观CSAR搭配FP4已经能把全局KV压缩到V4时期容量的四分之一左右而将SWAKV移出持久化存储后缓存规模几乎直接再减一半回看V4的设计持久化KV的占用空间大概一半来自SWA另一半来自全局KVV4.1首先通过CSAR和FP4把这占一半" } ], "id": 431, "review_annotation": "", "row_id": 431, "bounded_row_id": 431, "parent_row_ids": [ 431 ], "parent_row_id": 431, "packet_id": 32, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 882.0155384615385, "packet_end": 911.2035240384616, "raw_text": "缓存规模几乎直接再减一半", "selected_text": "缓存规模几乎直接再减一半", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 431 ], "packet_id": 32, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 900.5155384615385, "end": 902.0555384615385, "text": "回看V4的设计", "raw_asr_text": "回看V4的设计", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 32, "words": [ { "word": "回", "start": 900.5155384615385, "end": 900.8955384615385, "probability": 0.96728515625 }, { "word": "看", "start": 900.8955384615385, "end": 901.1355384615385, "probability": 0.9990234375 }, { "word": "V", "start": 901.1355384615385, "end": 901.3755384615386, "probability": 0.97119140625 }, { "word": "4", "start": 901.3755384615386, "end": 901.5955384615386, "probability": 0.9970703125 }, { "word": "的", "start": 901.5955384615386, "end": 901.7155384615386, "probability": 0.99169921875 }, { "word": "设", "start": 901.7155384615386, "end": 901.8955384615385, "probability": 1.0 }, { "word": "计", "start": 901.8955384615385, "end": 902.0555384615385, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10354346613730153, "compression_ratio": 1.3265993265993266, "no_speech_prob": 1.9754167052132843e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "与此同时全局KV可以存储在固态硬盘这类持久化存储中这部分的持久化KV缩减幅度非常可观CSAR搭配FP4已经能把全局KV压缩到V4时期容量的四分之一左右而将SWAKV移出持久化存储后缓存规模几乎直接再减一半回看V4的设计持久化KV的占用空间大概一半来自SWA另一半来自全局KVV4.1首先通过CSAR和FP4把这占一半" } ], "id": 432, "review_annotation": "", "row_id": 432, "bounded_row_id": 432, "parent_row_ids": [ 432 ], "parent_row_id": 432, "packet_id": 32, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 882.0155384615385, "packet_end": 911.2035240384616, "raw_text": "回看V4的设计", "selected_text": "回看V4的设计", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 432 ], "packet_id": 32, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 902.0555384615385, "end": 905.3755384615386, "text": "持久化KV的占用空间大概一半来自SWA", "raw_asr_text": "持久化KV的占用空间大概一半来自SWA", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 32, "words": [ { "word": "持", "start": 902.0555384615385, "end": 902.3155384615385, "probability": 0.962890625 }, { "word": "久", "start": 902.3155384615385, "end": 902.4755384615386, "probability": 1.0 }, { "word": "化", "start": 902.4755384615386, "end": 902.6355384615385, "probability": 0.9970703125 }, { "word": "K", "start": 902.6355384615385, "end": 902.8155384615385, "probability": 0.99755859375 }, { "word": "V", "start": 902.8155384615385, "end": 902.9355384615385, "probability": 1.0 }, { "word": "的", "start": 902.9355384615385, "end": 903.0555384615385, "probability": 0.97314453125 }, { "word": "占", "start": 903.0555384615385, "end": 903.1755384615385, "probability": 0.9951171875 }, { "word": "用", "start": 903.1755384615385, "end": 903.3355384615386, "probability": 0.99853515625 }, { "word": "空", "start": 903.3355384615386, "end": 903.4955384615386, "probability": 0.9990234375 }, { "word": "间", "start": 903.4955384615386, "end": 903.6955384615385, "probability": 0.99951171875 }, { "word": "大概", "start": 903.6955384615385, "end": 903.9955384615386, "probability": 0.98583984375 }, { "word": "一", "start": 903.9955384615386, "end": 904.1955384615385, "probability": 0.99462890625 }, { "word": "半", "start": 904.1955384615385, "end": 904.3155384615385, "probability": 0.99755859375 }, { "word": "来", "start": 904.3155384615385, "end": 904.5555384615385, "probability": 0.99755859375 }, { "word": "自", "start": 904.5555384615385, "end": 904.7355384615386, "probability": 0.99951171875 }, { "word": "S", "start": 904.7355384615386, "end": 904.9355384615385, "probability": 0.94677734375 }, { "word": "WA", "start": 904.9355384615385, "end": 905.3755384615386, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.10354346613730153, "compression_ratio": 1.3265993265993266, "no_speech_prob": 1.9754167052132843e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "与此同时全局KV可以存储在固态硬盘这类持久化存储中这部分的持久化KV缩减幅度非常可观CSAR搭配FP4已经能把全局KV压缩到V4时期容量的四分之一左右而将SWAKV移出持久化存储后缓存规模几乎直接再减一半回看V4的设计持久化KV的占用空间大概一半来自SWA另一半来自全局KVV4.1首先通过CSAR和FP4把这占一半" } ], "id": 433, "review_annotation": "", "row_id": 433, "bounded_row_id": 433, "parent_row_ids": [ 433 ], "parent_row_id": 433, "packet_id": 32, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 882.0155384615385, "packet_end": 911.2035240384616, "raw_text": "持久化KV的占用空间大概一半来自SWA", "selected_text": "持久化KV的占用空间大概一半来自SWA", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 433 ], "packet_id": 32, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 905.3755384615386, "end": 907.3355384615386, "text": "另一半来自全局KV", "raw_asr_text": "另一半来自全局KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 32, "words": [ { "word": "另", "start": 905.3755384615386, "end": 905.7955384615385, "probability": 0.927734375 }, { "word": "一", "start": 905.7955384615385, "end": 905.9155384615385, "probability": 0.99951171875 }, { "word": "半", "start": 905.9155384615385, "end": 906.0355384615385, "probability": 0.998046875 }, { "word": "来", "start": 906.0355384615385, "end": 906.2555384615385, "probability": 0.99951171875 }, { "word": "自", "start": 906.2555384615385, "end": 906.4355384615385, "probability": 0.99853515625 }, { "word": "全", "start": 906.4355384615385, "end": 906.7555384615385, "probability": 0.99853515625 }, { "word": "局", "start": 906.7555384615385, "end": 906.9555384615386, "probability": 1.0 }, { "word": "K", "start": 906.9555384615386, "end": 907.1355384615385, "probability": 0.99658203125 }, { "word": "V", "start": 907.1355384615385, "end": 907.3355384615386, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.10354346613730153, "compression_ratio": 1.3265993265993266, "no_speech_prob": 1.9754167052132843e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "与此同时全局KV可以存储在固态硬盘这类持久化存储中这部分的持久化KV缩减幅度非常可观CSAR搭配FP4已经能把全局KV压缩到V4时期容量的四分之一左右而将SWAKV移出持久化存储后缓存规模几乎直接再减一半回看V4的设计持久化KV的占用空间大概一半来自SWA另一半来自全局KVV4.1首先通过CSAR和FP4把这占一半" } ], "id": 434, "review_annotation": "", "row_id": 434, "bounded_row_id": 434, "parent_row_ids": [ 434 ], "parent_row_id": 434, "packet_id": 32, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 882.0155384615385, "packet_end": 911.2035240384616, "raw_text": "另一半来自全局KV", "selected_text": "另一半来自全局KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 434 ], "packet_id": 32, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 907.3355384615386, "end": 910.4035240384617, "text": "V4.1首先通过CSAR和FP4把这占一半", "raw_asr_text": "V4.1首先通过CSAR和FP4把这占一半", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 32, "words": [ { "word": "V", "start": 907.3355384615386, "end": 907.5955384615386, "probability": 0.9150390625 }, { "word": "4.", "start": 907.5955384615386, "end": 907.7755384615385, "probability": 0.986328125 }, { "word": "1", "start": 907.9155384615385, "end": 908.0955384615386, "probability": 0.998046875 }, { "word": "首先", "start": 908.0955384615386, "end": 908.5555384615385, "probability": 0.966796875 }, { "word": "通", "start": 908.5555384615385, "end": 908.7355384615386, "probability": 0.99853515625 }, { "word": "过", "start": 908.7355384615386, "end": 908.9355384615385, "probability": 0.99951171875 }, { "word": "CS", "start": 908.9355384615385, "end": 909.1955384615385, "probability": 0.64208984375 }, { "word": "AR", "start": 909.1955384615385, "end": 909.4955384615386, "probability": 0.99609375 }, { "word": "和", "start": 909.4955384615386, "end": 909.7755384615385, "probability": 0.97607421875 }, { "word": "FP", "start": 909.7755384615385, "end": 910.1555384615385, "probability": 0.986328125 }, { "word": "4", "start": 910.1555384615385, "end": 910.4155384615385, "probability": 0.99462890625 }, { "word": "把", "start": 910.4155384615385, "end": 910.6355384615385, "probability": 0.95849609375 }, { "word": "这", "start": 910.6355384615385, "end": 910.7955384615385, "probability": 0.97802734375 }, { "word": "占", "start": 910.7955384615385, "end": 910.9955384615386, "probability": 0.98876953125 }, { "word": "一", "start": 910.9955384615386, "end": 911.1755384615385, "probability": 0.9091796875 }, { "word": "半", "start": 911.1755384615385, "end": 911.1755384615385, "probability": 0.9443359375 } ], "asr_metrics": { "avg_logprob": -0.10354346613730153, "compression_ratio": 1.3265993265993266, "no_speech_prob": 1.9754167052132843e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "与此同时全局KV可以存储在固态硬盘这类持久化存储中这部分的持久化KV缩减幅度非常可观CSAR搭配FP4已经能把全局KV压缩到V4时期容量的四分之一左右而将SWAKV移出持久化存储后缓存规模几乎直接再减一半回看V4的设计持久化KV的占用空间大概一半来自SWA另一半来自全局KVV4.1首先通过CSAR和FP4把这占一半" } ], "id": 435, "review_annotation": "", "row_id": 435, "bounded_row_id": 435, "parent_row_ids": [ 435 ], "parent_row_id": 435, "packet_id": 32, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 882.0155384615385, "packet_end": 911.2035240384616, "raw_text": "V4.1首先通过CSAR和FP4把这占一半", "selected_text": "V4.1首先通过CSAR和FP4把这占一半", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 435 ], "packet_id": 32, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 910.4035240384617, "end": 912.2835240384617, "text": "和FP4把这占一半的全局KV", "raw_asr_text": "和FP4把这占一半的全局KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 33, "words": [ { "word": "和", "start": 909.6035240384617, "end": 909.7635240384617, "probability": 0.423828125 }, { "word": "FP", "start": 909.7635240384617, "end": 910.1235240384617, "probability": 0.810546875 }, { "word": "4", "start": 910.1235240384617, "end": 910.4035240384617, "probability": 0.9951171875 }, { "word": "把", "start": 910.4035240384617, "end": 910.6435240384617, "probability": 0.81103515625 }, { "word": "这", "start": 910.6435240384617, "end": 910.8035240384618, "probability": 0.8955078125 }, { "word": "占", "start": 910.8035240384618, "end": 910.9435240384618, "probability": 0.9873046875 }, { "word": "一", "start": 910.9435240384618, "end": 911.2035240384618, "probability": 0.99609375 }, { "word": "半", "start": 911.2035240384618, "end": 911.3835240384617, "probability": 0.9990234375 }, { "word": "的", "start": 911.3835240384617, "end": 911.5835240384617, "probability": 0.984375 }, { "word": "全", "start": 911.5835240384617, "end": 911.7835240384617, "probability": 0.98828125 }, { "word": "局", "start": 911.7835240384617, "end": 911.8835240384617, "probability": 0.99853515625 }, { "word": "K", "start": 911.8835240384617, "end": 912.0835240384617, "probability": 0.8232421875 }, { "word": "V", "start": 912.0835240384617, "end": 912.2835240384617, "probability": 0.9970703125 } ], "asr_metrics": { "avg_logprob": -0.11995403657709887, "compression_ratio": 1.1887905604719764, "no_speech_prob": 2.6069696748565185e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和FP4把这占一半的全局KV压缩到原大小的四分之一左右这就使得全局KV本身的占用降至每Token近890字节在HBM存储上直接节省了75%的空间之后再把占一半的SWA KV移出持久化存储固态硬盘存储空间最多可节省87.5%这也让这个总参数量达5520亿的模型在实际运行时呈现出一种很奇特的不对称性因此V4.1仍然属于MOE模型" } ], "id": 436, "review_annotation": "", "row_id": 436, "bounded_row_id": 436, "parent_row_ids": [ 436 ], "parent_row_id": 436, "packet_id": 33, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 909.6035240384617, "packet_end": 938.7915096153847, "raw_text": "和FP4把这占一半的全局KV", "selected_text": "和FP4把这占一半的全局KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 436 ], "packet_id": 33, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 912.2835240384617, "end": 914.6035240384617, "text": "压缩到原大小的四分之一左右", "raw_asr_text": "压缩到原大小的四分之一左右", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 33, "words": [ { "word": "压", "start": 912.2835240384617, "end": 912.6035240384617, "probability": 0.982421875 }, { "word": "缩", "start": 912.6035240384617, "end": 912.8235240384618, "probability": 0.998046875 }, { "word": "到", "start": 912.8235240384618, "end": 912.9835240384617, "probability": 0.99365234375 }, { "word": "原", "start": 912.9835240384617, "end": 913.2635240384617, "probability": 0.87109375 }, { "word": "大", "start": 913.2635240384617, "end": 913.4435240384618, "probability": 0.99755859375 }, { "word": "小", "start": 913.4435240384618, "end": 913.6435240384617, "probability": 0.9990234375 }, { "word": "的", "start": 913.6435240384617, "end": 913.7835240384617, "probability": 0.99609375 }, { "word": "四", "start": 913.7835240384617, "end": 913.9235240384618, "probability": 0.442626953125 }, { "word": "分", "start": 913.9235240384618, "end": 914.0435240384618, "probability": 0.99853515625 }, { "word": "之", "start": 914.0435240384618, "end": 914.1835240384618, "probability": 0.9970703125 }, { "word": "一", "start": 914.1835240384618, "end": 914.3035240384618, "probability": 0.98779296875 }, { "word": "左右", "start": 914.3035240384618, "end": 914.6035240384617, "probability": 0.98486328125 } ], "asr_metrics": { "avg_logprob": -0.11995403657709887, "compression_ratio": 1.1887905604719764, "no_speech_prob": 2.6069696748565185e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和FP4把这占一半的全局KV压缩到原大小的四分之一左右这就使得全局KV本身的占用降至每Token近890字节在HBM存储上直接节省了75%的空间之后再把占一半的SWA KV移出持久化存储固态硬盘存储空间最多可节省87.5%这也让这个总参数量达5520亿的模型在实际运行时呈现出一种很奇特的不对称性因此V4.1仍然属于MOE模型" } ], "id": 437, "review_annotation": "", "row_id": 437, "bounded_row_id": 437, "parent_row_ids": [ 437 ], "parent_row_id": 437, "packet_id": 33, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 909.6035240384617, "packet_end": 938.7915096153847, "raw_text": "压缩到原大小的四分之一左右", "selected_text": "压缩到原大小的四分之一左右", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 437 ], "packet_id": 33, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 914.6035240384617, "end": 917.1235240384617, "text": "这就使得全局KV本身的占用", "raw_asr_text": "这就使得全局KV本身的占用", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 33, "words": [ { "word": "这", "start": 914.6035240384617, "end": 914.9435240384618, "probability": 0.9794921875 }, { "word": "就", "start": 914.9435240384618, "end": 915.0835240384617, "probability": 0.9970703125 }, { "word": "使", "start": 915.0835240384617, "end": 915.2835240384617, "probability": 0.9990234375 }, { "word": "得", "start": 915.2835240384617, "end": 915.4635240384617, "probability": 0.99951171875 }, { "word": "全", "start": 915.4635240384617, "end": 915.7235240384617, "probability": 0.99609375 }, { "word": "局", "start": 915.7235240384617, "end": 915.8835240384617, "probability": 1.0 }, { "word": "K", "start": 915.8835240384617, "end": 916.0635240384618, "probability": 0.99462890625 }, { "word": "V", "start": 916.0635240384618, "end": 916.2435240384617, "probability": 0.99951171875 }, { "word": "本", "start": 916.2435240384617, "end": 916.4835240384617, "probability": 0.9775390625 }, { "word": "身", "start": 916.4835240384617, "end": 916.6635240384617, "probability": 1.0 }, { "word": "的", "start": 916.6635240384617, "end": 916.7835240384617, "probability": 0.99267578125 }, { "word": "占", "start": 916.7835240384617, "end": 916.9035240384617, "probability": 0.998046875 }, { "word": "用", "start": 916.9035240384617, "end": 917.1235240384617, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11995403657709887, "compression_ratio": 1.1887905604719764, "no_speech_prob": 2.6069696748565185e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和FP4把这占一半的全局KV压缩到原大小的四分之一左右这就使得全局KV本身的占用降至每Token近890字节在HBM存储上直接节省了75%的空间之后再把占一半的SWA KV移出持久化存储固态硬盘存储空间最多可节省87.5%这也让这个总参数量达5520亿的模型在实际运行时呈现出一种很奇特的不对称性因此V4.1仍然属于MOE模型" } ], "id": 438, "review_annotation": "", "row_id": 438, "bounded_row_id": 438, "parent_row_ids": [ 438 ], "parent_row_id": 438, "packet_id": 33, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 909.6035240384617, "packet_end": 938.7915096153847, "raw_text": "这就使得全局KV本身的占用", "selected_text": "这就使得全局KV本身的占用", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 438 ], "packet_id": 33, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 917.1235240384617, "end": 919.4235240384618, "text": "降至每Token近890字节", "raw_asr_text": "降至每Token近890字节", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 33, "words": [ { "word": "降", "start": 917.1235240384617, "end": 917.3635240384617, "probability": 0.9814453125 }, { "word": "至", "start": 917.3635240384617, "end": 917.5435240384618, "probability": 0.88427734375 }, { "word": "每", "start": 917.5435240384618, "end": 917.6635240384617, "probability": 0.9970703125 }, { "word": "T", "start": 917.6635240384617, "end": 917.8235240384618, "probability": 0.6162109375 }, { "word": "oken", "start": 917.8235240384618, "end": 918.0235240384617, "probability": 0.95556640625 }, { "word": "近", "start": 918.0235240384617, "end": 918.3035240384618, "probability": 0.51513671875 }, { "word": "8", "start": 918.3035240384618, "end": 918.5835240384617, "probability": 0.97119140625 }, { "word": "90", "start": 918.5835240384617, "end": 919.0635240384618, "probability": 0.9921875 }, { "word": "字", "start": 919.0635240384618, "end": 919.2435240384617, "probability": 0.92333984375 }, { "word": "节", "start": 919.2435240384617, "end": 919.4235240384618, "probability": 0.9892578125 } ], "asr_metrics": { "avg_logprob": -0.11995403657709887, "compression_ratio": 1.1887905604719764, "no_speech_prob": 2.6069696748565185e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和FP4把这占一半的全局KV压缩到原大小的四分之一左右这就使得全局KV本身的占用降至每Token近890字节在HBM存储上直接节省了75%的空间之后再把占一半的SWA KV移出持久化存储固态硬盘存储空间最多可节省87.5%这也让这个总参数量达5520亿的模型在实际运行时呈现出一种很奇特的不对称性因此V4.1仍然属于MOE模型" } ], "id": 439, "review_annotation": "", "row_id": 439, "bounded_row_id": 439, "parent_row_ids": [ 439 ], "parent_row_id": 439, "packet_id": 33, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 909.6035240384617, "packet_end": 938.7915096153847, "raw_text": "降至每Token近890字节", "selected_text": "降至每Token近890字节", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 439 ], "packet_id": 33, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 919.4235240384618, "end": 921.0435240384618, "text": "在HBM存储上", "raw_asr_text": "在HBM存储上", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 33, "words": [ { "word": "在", "start": 919.4235240384618, "end": 919.9235240384618, "probability": 0.98291015625 }, { "word": "H", "start": 919.9235240384618, "end": 920.1835240384618, "probability": 0.99560546875 }, { "word": "BM", "start": 920.1835240384618, "end": 920.4635240384617, "probability": 0.94873046875 }, { "word": "存", "start": 920.4635240384617, "end": 920.7035240384618, "probability": 0.99853515625 }, { "word": "储", "start": 920.7035240384618, "end": 920.8635240384617, "probability": 0.998046875 }, { "word": "上", "start": 920.8635240384617, "end": 921.0435240384618, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.11995403657709887, "compression_ratio": 1.1887905604719764, "no_speech_prob": 2.6069696748565185e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和FP4把这占一半的全局KV压缩到原大小的四分之一左右这就使得全局KV本身的占用降至每Token近890字节在HBM存储上直接节省了75%的空间之后再把占一半的SWA KV移出持久化存储固态硬盘存储空间最多可节省87.5%这也让这个总参数量达5520亿的模型在实际运行时呈现出一种很奇特的不对称性因此V4.1仍然属于MOE模型" } ], "id": 440, "review_annotation": "", "row_id": 440, "bounded_row_id": 440, "parent_row_ids": [ 440 ], "parent_row_id": 440, "packet_id": 33, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 909.6035240384617, "packet_end": 938.7915096153847, "raw_text": "在HBM存储上", "selected_text": "在HBM存储上", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 440 ], "packet_id": 33, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 921.0435240384618, "end": 923.2035240384618, "text": "直接节省了75%的空间", "raw_asr_text": "直接节省了75%的空间", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 33, "words": [ { "word": "直接", "start": 921.0435240384618, "end": 921.4035240384617, "probability": 0.98974609375 }, { "word": "节", "start": 921.4035240384617, "end": 921.6235240384617, "probability": 0.98486328125 }, { "word": "省", "start": 921.6235240384617, "end": 921.8435240384617, "probability": 0.99951171875 }, { "word": "了", "start": 921.8435240384617, "end": 921.9835240384617, "probability": 0.998046875 }, { "word": "75", "start": 921.9835240384617, "end": 922.5435240384618, "probability": 0.98681640625 }, { "word": "%", "start": 922.5435240384618, "end": 922.5435240384618, "probability": 0.99853515625 }, { "word": "的", "start": 922.5435240384618, "end": 922.8835240384617, "probability": 0.98486328125 }, { "word": "空", "start": 922.8835240384617, "end": 923.0235240384617, "probability": 0.9990234375 }, { "word": "间", "start": 923.0235240384617, "end": 923.2035240384618, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11995403657709887, "compression_ratio": 1.1887905604719764, "no_speech_prob": 2.6069696748565185e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和FP4把这占一半的全局KV压缩到原大小的四分之一左右这就使得全局KV本身的占用降至每Token近890字节在HBM存储上直接节省了75%的空间之后再把占一半的SWA KV移出持久化存储固态硬盘存储空间最多可节省87.5%这也让这个总参数量达5520亿的模型在实际运行时呈现出一种很奇特的不对称性因此V4.1仍然属于MOE模型" } ], "id": 441, "review_annotation": "", "row_id": 441, "bounded_row_id": 441, "parent_row_ids": [ 441 ], "parent_row_id": 441, "packet_id": 33, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 909.6035240384617, "packet_end": 938.7915096153847, "raw_text": "直接节省了75%的空间", "selected_text": "直接节省了75%的空间", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 441 ], "packet_id": 33, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 923.2035240384618, "end": 925.5035240384617, "text": "之后再把占一半的SWA KV", "raw_asr_text": "之后再把占一半的SWA KV", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 33, "words": [ { "word": "之", "start": 923.2035240384618, "end": 923.4635240384617, "probability": 0.98486328125 }, { "word": "后", "start": 923.4635240384617, "end": 923.5835240384617, "probability": 1.0 }, { "word": "再", "start": 923.5835240384617, "end": 923.7235240384617, "probability": 0.98583984375 }, { "word": "把", "start": 923.7235240384617, "end": 923.8835240384617, "probability": 0.97998046875 }, { "word": "占", "start": 923.8835240384617, "end": 924.0835240384617, "probability": 0.9990234375 }, { "word": "一", "start": 924.0835240384617, "end": 924.2435240384617, "probability": 0.9990234375 }, { "word": "半", "start": 924.2435240384617, "end": 924.3635240384617, "probability": 0.99951171875 }, { "word": "的", "start": 924.3635240384617, "end": 924.5635240384618, "probability": 0.9970703125 }, { "word": "S", "start": 924.5635240384618, "end": 924.7635240384617, "probability": 0.9892578125 }, { "word": "WA", "start": 924.7635240384617, "end": 925.1235240384617, "probability": 0.82470703125 }, { "word": " K", "start": 925.1235240384617, "end": 925.3035240384618, "probability": 0.40283203125 }, { "word": "V", "start": 925.3035240384618, "end": 925.5035240384617, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.11995403657709887, "compression_ratio": 1.1887905604719764, "no_speech_prob": 2.6069696748565185e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和FP4把这占一半的全局KV压缩到原大小的四分之一左右这就使得全局KV本身的占用降至每Token近890字节在HBM存储上直接节省了75%的空间之后再把占一半的SWA KV移出持久化存储固态硬盘存储空间最多可节省87.5%这也让这个总参数量达5520亿的模型在实际运行时呈现出一种很奇特的不对称性因此V4.1仍然属于MOE模型" } ], "id": 442, "review_annotation": "", "row_id": 442, "bounded_row_id": 442, "parent_row_ids": [ 442 ], "parent_row_id": 442, "packet_id": 33, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 909.6035240384617, "packet_end": 938.7915096153847, "raw_text": "之后再把占一半的SWA KV", "selected_text": "之后再把占一半的SWA KV", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 442 ], "packet_id": 33, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 925.5035240384617, "end": 928.1435240384617, "text": "移出持久化存储固态硬盘存储空间", "raw_asr_text": "移出持久化存储固态硬盘存储空间", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 33, "words": [ { "word": "移", "start": 925.5035240384617, "end": 925.7635240384617, "probability": 0.99609375 }, { "word": "出", "start": 925.7635240384617, "end": 925.9435240384618, "probability": 0.99462890625 }, { "word": "持", "start": 925.9435240384618, "end": 926.1835240384618, "probability": 0.99853515625 }, { "word": "久", "start": 926.1835240384618, "end": 926.3635240384617, "probability": 1.0 }, { "word": "化", "start": 926.3635240384617, "end": 926.5235240384617, "probability": 0.99658203125 }, { "word": "存", "start": 926.5235240384617, "end": 926.7235240384617, "probability": 0.98388671875 }, { "word": "储", "start": 926.7235240384617, "end": 926.8635240384617, "probability": 0.99951171875 }, { "word": "固", "start": 926.8635240384617, "end": 927.0235240384617, "probability": 0.998046875 }, { "word": "态", "start": 927.0235240384617, "end": 927.1635240384617, "probability": 0.9921875 }, { "word": "硬", "start": 927.1635240384617, "end": 927.2835240384617, "probability": 0.9990234375 }, { "word": "盘", "start": 927.2835240384617, "end": 927.4635240384617, "probability": 0.9814453125 }, { "word": "存", "start": 927.4635240384617, "end": 927.6435240384617, "probability": 0.98486328125 }, { "word": "储", "start": 927.6435240384617, "end": 927.8035240384618, "probability": 1.0 }, { "word": "空", "start": 927.8035240384618, "end": 927.9835240384617, "probability": 0.99267578125 }, { "word": "间", "start": 927.9835240384617, "end": 928.1435240384617, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.11995403657709887, "compression_ratio": 1.1887905604719764, "no_speech_prob": 2.6069696748565185e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和FP4把这占一半的全局KV压缩到原大小的四分之一左右这就使得全局KV本身的占用降至每Token近890字节在HBM存储上直接节省了75%的空间之后再把占一半的SWA KV移出持久化存储固态硬盘存储空间最多可节省87.5%这也让这个总参数量达5520亿的模型在实际运行时呈现出一种很奇特的不对称性因此V4.1仍然属于MOE模型" } ], "id": 443, "review_annotation": "", "row_id": 443, "bounded_row_id": 443, "parent_row_ids": [ 443 ], "parent_row_id": 443, "packet_id": 33, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 909.6035240384617, "packet_end": 938.7915096153847, "raw_text": "移出持久化存储固态硬盘存储空间", "selected_text": "移出持久化存储固态硬盘存储空间", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 443 ], "packet_id": 33, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 928.1435240384617, "end": 930.3235240384618, "text": "最多可节省87.5%", "raw_asr_text": "最多可节省87.5%", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 33, "words": [ { "word": "最", "start": 928.1435240384617, "end": 928.3435240384617, "probability": 0.9912109375 }, { "word": "多", "start": 928.3435240384617, "end": 928.4835240384617, "probability": 0.9990234375 }, { "word": "可", "start": 928.4835240384617, "end": 928.6035240384617, "probability": 0.99072265625 }, { "word": "节", "start": 928.6035240384617, "end": 928.7235240384617, "probability": 0.98779296875 }, { "word": "省", "start": 928.7235240384617, "end": 928.8835240384617, "probability": 1.0 }, { "word": "87.", "start": 928.8835240384617, "end": 929.5835240384617, "probability": 0.984375 }, { "word": "5", "start": 929.9435240384618, "end": 930.3035240384618, "probability": 0.89697265625 }, { "word": "%", "start": 930.3035240384618, "end": 930.3235240384618, "probability": 0.9951171875 } ], "asr_metrics": { "avg_logprob": -0.11995403657709887, "compression_ratio": 1.1887905604719764, "no_speech_prob": 2.6069696748565185e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和FP4把这占一半的全局KV压缩到原大小的四分之一左右这就使得全局KV本身的占用降至每Token近890字节在HBM存储上直接节省了75%的空间之后再把占一半的SWA KV移出持久化存储固态硬盘存储空间最多可节省87.5%这也让这个总参数量达5520亿的模型在实际运行时呈现出一种很奇特的不对称性因此V4.1仍然属于MOE模型" } ], "id": 444, "review_annotation": "", "row_id": 444, "bounded_row_id": 444, "parent_row_ids": [ 444 ], "parent_row_id": 444, "packet_id": 33, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 909.6035240384617, "packet_end": 938.7915096153847, "raw_text": "最多可节省87.5%", "selected_text": "最多可节省87.5%", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 444 ], "packet_id": 33, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 930.3235240384618, "end": 933.1835240384618, "text": "这也让这个总参数量达5520亿的模型", "raw_asr_text": "这也让这个总参数量达5520亿的模型", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 33, "words": [ { "word": "这", "start": 930.3235240384618, "end": 930.5235240384617, "probability": 0.97900390625 }, { "word": "也", "start": 930.5235240384617, "end": 930.6435240384617, "probability": 0.99853515625 }, { "word": "让", "start": 930.6435240384617, "end": 930.7835240384617, "probability": 0.99853515625 }, { "word": "这个", "start": 930.7835240384617, "end": 930.9435240384618, "probability": 0.89208984375 }, { "word": "总", "start": 930.9435240384618, "end": 931.1435240384617, "probability": 0.99853515625 }, { "word": "参", "start": 931.1435240384617, "end": 931.3035240384618, "probability": 0.85986328125 }, { "word": "数", "start": 931.3035240384618, "end": 931.4435240384618, "probability": 1.0 }, { "word": "量", "start": 931.4435240384618, "end": 931.5635240384618, "probability": 0.99462890625 }, { "word": "达", "start": 931.5635240384618, "end": 931.7635240384617, "probability": 0.9931640625 }, { "word": "55", "start": 931.7635240384617, "end": 932.0435240384618, "probability": 0.9248046875 }, { "word": "20", "start": 932.0435240384618, "end": 932.6235240384617, "probability": 0.9990234375 }, { "word": "亿", "start": 932.6235240384617, "end": 932.7435240384617, "probability": 0.99609375 }, { "word": "的", "start": 932.7435240384617, "end": 932.8435240384617, "probability": 0.998046875 }, { "word": "模", "start": 932.8435240384617, "end": 932.9835240384617, "probability": 0.9990234375 }, { "word": "型", "start": 932.9835240384617, "end": 933.1835240384618, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11995403657709887, "compression_ratio": 1.1887905604719764, "no_speech_prob": 2.6069696748565185e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和FP4把这占一半的全局KV压缩到原大小的四分之一左右这就使得全局KV本身的占用降至每Token近890字节在HBM存储上直接节省了75%的空间之后再把占一半的SWA KV移出持久化存储固态硬盘存储空间最多可节省87.5%这也让这个总参数量达5520亿的模型在实际运行时呈现出一种很奇特的不对称性因此V4.1仍然属于MOE模型" } ], "id": 445, "review_annotation": "", "row_id": 445, "bounded_row_id": 445, "parent_row_ids": [ 445 ], "parent_row_id": 445, "packet_id": 33, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 909.6035240384617, "packet_end": 938.7915096153847, "raw_text": "这也让这个总参数量达5520亿的模型", "selected_text": "这也让这个总参数量达5520亿的模型", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 445 ], "packet_id": 33, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 933.1835240384618, "end": 934.1635240384617, "text": "在实际运行时", "raw_asr_text": "在实际运行时", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 33, "words": [ { "word": "在", "start": 933.1835240384618, "end": 933.4035240384617, "probability": 0.986328125 }, { "word": "实", "start": 933.4035240384617, "end": 933.5635240384618, "probability": 0.99755859375 }, { "word": "际", "start": 933.5635240384618, "end": 933.6835240384618, "probability": 0.9990234375 }, { "word": "运", "start": 933.6835240384618, "end": 933.8235240384618, "probability": 1.0 }, { "word": "行", "start": 933.8235240384618, "end": 933.9835240384617, "probability": 0.99658203125 }, { "word": "时", "start": 933.9835240384617, "end": 934.1635240384617, "probability": 0.9951171875 } ], "asr_metrics": { "avg_logprob": -0.11995403657709887, "compression_ratio": 1.1887905604719764, "no_speech_prob": 2.6069696748565185e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和FP4把这占一半的全局KV压缩到原大小的四分之一左右这就使得全局KV本身的占用降至每Token近890字节在HBM存储上直接节省了75%的空间之后再把占一半的SWA KV移出持久化存储固态硬盘存储空间最多可节省87.5%这也让这个总参数量达5520亿的模型在实际运行时呈现出一种很奇特的不对称性因此V4.1仍然属于MOE模型" } ], "id": 446, "review_annotation": "", "row_id": 446, "bounded_row_id": 446, "parent_row_ids": [ 446 ], "parent_row_id": 446, "packet_id": 33, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 909.6035240384617, "packet_end": 938.7915096153847, "raw_text": "在实际运行时", "selected_text": "在实际运行时", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 446 ], "packet_id": 33, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 934.1635240384617, "end": 936.2635240384617, "text": "呈现出一种很奇特的不对称性", "raw_asr_text": "呈现出一种很奇特的不对称性", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 33, "words": [ { "word": "呈", "start": 934.1635240384617, "end": 934.4035240384617, "probability": 0.99462890625 }, { "word": "现", "start": 934.4035240384617, "end": 934.5435240384618, "probability": 0.9990234375 }, { "word": "出", "start": 934.5435240384618, "end": 934.6835240384618, "probability": 0.998046875 }, { "word": "一", "start": 934.6835240384618, "end": 934.8035240384618, "probability": 0.994140625 }, { "word": "种", "start": 934.8035240384618, "end": 934.9235240384618, "probability": 0.99951171875 }, { "word": "很", "start": 934.9235240384618, "end": 935.1835240384618, "probability": 0.994140625 }, { "word": "奇", "start": 935.1835240384618, "end": 935.3835240384617, "probability": 0.998046875 }, { "word": "特", "start": 935.3835240384617, "end": 935.5435240384618, "probability": 0.9990234375 }, { "word": "的", "start": 935.5435240384618, "end": 935.6835240384618, "probability": 0.99853515625 }, { "word": "不对", "start": 935.6835240384618, "end": 935.9235240384618, "probability": 0.99169921875 }, { "word": "称", "start": 935.9235240384618, "end": 936.1035240384617, "probability": 0.99609375 }, { "word": "性", "start": 936.1035240384617, "end": 936.2635240384617, "probability": 0.99267578125 } ], "asr_metrics": { "avg_logprob": -0.11995403657709887, "compression_ratio": 1.1887905604719764, "no_speech_prob": 2.6069696748565185e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和FP4把这占一半的全局KV压缩到原大小的四分之一左右这就使得全局KV本身的占用降至每Token近890字节在HBM存储上直接节省了75%的空间之后再把占一半的SWA KV移出持久化存储固态硬盘存储空间最多可节省87.5%这也让这个总参数量达5520亿的模型在实际运行时呈现出一种很奇特的不对称性因此V4.1仍然属于MOE模型" } ], "id": 447, "review_annotation": "", "row_id": 447, "bounded_row_id": 447, "parent_row_ids": [ 447 ], "parent_row_id": 447, "packet_id": 33, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 909.6035240384617, "packet_end": 938.7915096153847, "raw_text": "呈现出一种很奇特的不对称性", "selected_text": "呈现出一种很奇特的不对称性", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 447 ], "packet_id": 33, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 936.2635240384617, "end": 937.9915096153848, "text": "因此V4.1仍然属于MOE模型", "raw_asr_text": "因此V4.1仍然属于MOE模型", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 33, "words": [ { "word": "因", "start": 936.2635240384617, "end": 936.5235240384617, "probability": 0.99267578125 }, { "word": "此", "start": 936.5235240384617, "end": 936.6835240384618, "probability": 1.0 }, { "word": "V", "start": 936.6835240384618, "end": 936.8835240384617, "probability": 0.931640625 }, { "word": "4.", "start": 936.8835240384617, "end": 937.0235240384617, "probability": 0.99658203125 }, { "word": "1", "start": 937.1235240384617, "end": 937.3235240384618, "probability": 0.9931640625 }, { "word": "仍", "start": 937.3235240384618, "end": 937.5835240384617, "probability": 0.9970703125 }, { "word": "然", "start": 937.5835240384617, "end": 937.7235240384617, "probability": 0.9990234375 }, { "word": "属", "start": 937.7235240384617, "end": 937.9035240384617, "probability": 0.99951171875 }, { "word": "于", "start": 937.9035240384617, "end": 938.0435240384618, "probability": 0.99853515625 }, { "word": "MO", "start": 938.0435240384618, "end": 938.2635240384617, "probability": 0.935546875 }, { "word": "E", "start": 938.2635240384617, "end": 938.4035240384617, "probability": 0.97509765625 }, { "word": "模", "start": 938.4035240384617, "end": 938.5435240384618, "probability": 0.99658203125 }, { "word": "型", "start": 938.5435240384618, "end": 938.7235240384617, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11995403657709887, "compression_ratio": 1.1887905604719764, "no_speech_prob": 2.6069696748565185e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "和FP4把这占一半的全局KV压缩到原大小的四分之一左右这就使得全局KV本身的占用降至每Token近890字节在HBM存储上直接节省了75%的空间之后再把占一半的SWA KV移出持久化存储固态硬盘存储空间最多可节省87.5%这也让这个总参数量达5520亿的模型在实际运行时呈现出一种很奇特的不对称性因此V4.1仍然属于MOE模型" } ], "id": 448, "review_annotation": "", "row_id": 448, "bounded_row_id": 448, "parent_row_ids": [ 448 ], "parent_row_id": 448, "packet_id": 33, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 909.6035240384617, "packet_end": 938.7915096153847, "raw_text": "因此V4.1仍然属于MOE模型", "selected_text": "因此V4.1仍然属于MOE模型", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 448 ], "packet_id": 33, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 938.7315096153848, "end": 942.6315096153849, "text": "显然它不会为每个Token都激活全部5520亿参数", "raw_asr_text": "显然它不会为每个Token都激活全部5520亿参数", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 34, "words": [ { "word": "显", "start": 938.7315096153848, "end": 938.9715096153848, "probability": 0.9482421875 }, { "word": "然", "start": 938.9715096153848, "end": 939.0915096153848, "probability": 0.99951171875 }, { "word": "它", "start": 939.0915096153848, "end": 939.2515096153847, "probability": 0.9140625 }, { "word": "不会", "start": 939.2515096153847, "end": 939.5115096153849, "probability": 0.99658203125 }, { "word": "为", "start": 939.5115096153849, "end": 939.7115096153848, "probability": 0.9921875 }, { "word": "每", "start": 939.7115096153848, "end": 939.9115096153848, "probability": 0.9970703125 }, { "word": "个", "start": 939.9115096153848, "end": 940.0515096153848, "probability": 0.99462890625 }, { "word": "T", "start": 940.0515096153848, "end": 940.1515096153848, "probability": 0.57470703125 }, { "word": "oken", "start": 940.1515096153848, "end": 940.2915096153848, "probability": 0.966796875 }, { "word": "都", "start": 940.2915096153848, "end": 940.4515096153848, "probability": 0.98974609375 }, { "word": "激", "start": 940.4515096153848, "end": 940.6715096153848, "probability": 0.9990234375 }, { "word": "活", "start": 940.6715096153848, "end": 940.8115096153848, "probability": 0.99609375 }, { "word": "全部", "start": 940.8115096153848, "end": 941.2115096153848, "probability": 0.9580078125 }, { "word": "55", "start": 941.2115096153848, "end": 941.5715096153848, "probability": 0.8916015625 }, { "word": "20", "start": 941.5715096153848, "end": 942.1315096153849, "probability": 0.99853515625 }, { "word": "亿", "start": 942.1315096153849, "end": 942.2715096153848, "probability": 0.9951171875 }, { "word": "参", "start": 942.2715096153848, "end": 942.4515096153848, "probability": 0.98779296875 }, { "word": "数", "start": 942.4515096153848, "end": 942.6315096153849, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.10110379634291361, "compression_ratio": 1.2724719101123596, "no_speech_prob": 6.470405461422857e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "1仍然属于MOE模型显然它不会为每个Token都激活全部5520亿参数但借助CED与填充阶段的输入Token只需经过20层编码器因此DeepSeq在这阶段仅激活约80亿个参数但模型一旦开始生成内容每个新Token都必须同时经过编码器和解码器激活参数量随之升至160亿所以即便整个模型的总规模几乎是VS Flash的两倍模型中负责大部分繁重输入读取工作的部分其实非常难" } ], "id": 449, "review_annotation": "", "row_id": 449, "bounded_row_id": 449, "parent_row_ids": [ 449 ], "parent_row_id": 449, "packet_id": 34, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 937.1915096153848, "packet_end": 966.3794951923077, "raw_text": "显然它不会为每个Token都激活全部5520亿参数", "selected_text": "显然它不会为每个Token都激活全部5520亿参数", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 449 ], "packet_id": 34, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 942.6315096153849, "end": 947.4115096153848, "text": "但借助CED与填充阶段的输入Token只需经过20层编码器", "raw_asr_text": "但借助CED与填充阶段的输入Token只需经过20层编码器", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 34, "words": [ { "word": "但", "start": 942.6315096153849, "end": 942.8915096153848, "probability": 0.98486328125 }, { "word": "借", "start": 942.8915096153848, "end": 943.1915096153848, "probability": 0.98291015625 }, { "word": "助", "start": 943.1915096153848, "end": 943.3515096153848, "probability": 0.99853515625 }, { "word": "C", "start": 943.3515096153848, "end": 943.5715096153848, "probability": 0.8583984375 }, { "word": "ED", "start": 943.5715096153848, "end": 943.7715096153848, "probability": 0.994140625 }, { "word": "与", "start": 943.7715096153848, "end": 944.2515096153847, "probability": 0.919921875 }, { "word": "填", "start": 944.2515096153847, "end": 944.4515096153848, "probability": 0.998046875 }, { "word": "充", "start": 944.4515096153848, "end": 944.6115096153848, "probability": 0.998046875 }, { "word": "阶", "start": 944.6115096153848, "end": 944.7715096153848, "probability": 0.994140625 }, { "word": "段", "start": 944.7715096153848, "end": 944.9115096153848, "probability": 1.0 }, { "word": "的", "start": 944.9115096153848, "end": 944.9915096153848, "probability": 0.99072265625 }, { "word": "输", "start": 944.9915096153848, "end": 945.1315096153849, "probability": 0.9990234375 }, { "word": "入", "start": 945.1315096153849, "end": 945.2315096153848, "probability": 0.99951171875 }, { "word": "T", "start": 945.2315096153848, "end": 945.3915096153848, "probability": 0.666015625 }, { "word": "oken", "start": 945.3915096153848, "end": 945.5715096153848, "probability": 0.93017578125 }, { "word": "只", "start": 945.5715096153848, "end": 945.8515096153848, "probability": 0.98291015625 }, { "word": "需", "start": 945.8515096153848, "end": 946.0115096153849, "probability": 0.97998046875 }, { "word": "经", "start": 946.0115096153849, "end": 946.1515096153848, "probability": 0.99853515625 }, { "word": "过", "start": 946.1515096153848, "end": 946.3515096153848, "probability": 0.99853515625 }, { "word": "20", "start": 946.3515096153848, "end": 946.7115096153848, "probability": 0.9853515625 }, { "word": "层", "start": 946.7115096153848, "end": 946.9315096153848, "probability": 0.998046875 }, { "word": "编", "start": 946.9315096153848, "end": 947.0915096153848, "probability": 0.99658203125 }, { "word": "码", "start": 947.0915096153848, "end": 947.2515096153847, "probability": 0.9990234375 }, { "word": "器", "start": 947.2515096153847, "end": 947.4115096153848, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.10110379634291361, "compression_ratio": 1.2724719101123596, "no_speech_prob": 6.470405461422857e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "1仍然属于MOE模型显然它不会为每个Token都激活全部5520亿参数但借助CED与填充阶段的输入Token只需经过20层编码器因此DeepSeq在这阶段仅激活约80亿个参数但模型一旦开始生成内容每个新Token都必须同时经过编码器和解码器激活参数量随之升至160亿所以即便整个模型的总规模几乎是VS Flash的两倍模型中负责大部分繁重输入读取工作的部分其实非常难" } ], "id": 450, "review_annotation": "", "row_id": 450, "bounded_row_id": 450, "parent_row_ids": [ 450 ], "parent_row_id": 450, "packet_id": 34, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 937.1915096153848, "packet_end": 966.3794951923077, "raw_text": "但借助CED与填充阶段的输入Token只需经过20层编码器", "selected_text": "但借助CED与填充阶段的输入Token只需经过20层编码器", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 450 ], "packet_id": 34, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 947.4115096153848, "end": 951.2715096153848, "text": "因此DeepSeq在这阶段仅激活约80亿个参数", "raw_asr_text": "因此DeepSeq在这阶段仅激活约80亿个参数", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 34, "words": [ { "word": "因", "start": 947.4115096153848, "end": 947.8915096153848, "probability": 0.9892578125 }, { "word": "此", "start": 947.8915096153848, "end": 948.0515096153848, "probability": 0.99951171875 }, { "word": "De", "start": 948.0515096153848, "end": 948.1715096153848, "probability": 0.93115234375 }, { "word": "ep", "start": 948.1715096153848, "end": 948.2715096153848, "probability": 0.94921875 }, { "word": "Se", "start": 948.2715096153848, "end": 948.3915096153848, "probability": 0.3564453125 }, { "word": "q", "start": 948.3915096153848, "end": 948.4915096153848, "probability": 0.537109375 }, { "word": "在", "start": 948.4915096153848, "end": 948.7115096153848, "probability": 0.99267578125 }, { "word": "这", "start": 948.7115096153848, "end": 948.9315096153848, "probability": 0.99609375 }, { "word": "阶", "start": 948.9315096153848, "end": 949.1315096153849, "probability": 0.73291015625 }, { "word": "段", "start": 949.1315096153849, "end": 949.3115096153848, "probability": 1.0 }, { "word": "仅", "start": 949.3115096153848, "end": 949.5715096153848, "probability": 0.9931640625 }, { "word": "激", "start": 949.5715096153848, "end": 949.7715096153848, "probability": 0.9990234375 }, { "word": "活", "start": 949.7715096153848, "end": 949.9315096153848, "probability": 0.99951171875 }, { "word": "约", "start": 949.9315096153848, "end": 950.0515096153848, "probability": 0.951171875 }, { "word": "80", "start": 950.0515096153848, "end": 950.5915096153848, "probability": 0.99609375 }, { "word": "亿", "start": 950.5915096153848, "end": 950.7915096153848, "probability": 1.0 }, { "word": "个", "start": 950.7915096153848, "end": 950.9115096153848, "probability": 0.68310546875 }, { "word": "参", "start": 950.9115096153848, "end": 951.0915096153848, "probability": 0.9990234375 }, { "word": "数", "start": 951.0915096153848, "end": 951.2715096153848, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10110379634291361, "compression_ratio": 1.2724719101123596, "no_speech_prob": 6.470405461422857e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "1仍然属于MOE模型显然它不会为每个Token都激活全部5520亿参数但借助CED与填充阶段的输入Token只需经过20层编码器因此DeepSeq在这阶段仅激活约80亿个参数但模型一旦开始生成内容每个新Token都必须同时经过编码器和解码器激活参数量随之升至160亿所以即便整个模型的总规模几乎是VS Flash的两倍模型中负责大部分繁重输入读取工作的部分其实非常难" } ], "id": 451, "review_annotation": "", "row_id": 451, "bounded_row_id": 451, "parent_row_ids": [ 451 ], "parent_row_id": 451, "packet_id": 34, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 937.1915096153848, "packet_end": 966.3794951923077, "raw_text": "因此DeepSeq在这阶段仅激活约80亿个参数", "selected_text": "因此DeepSeq在这阶段仅激活约80亿个参数", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 451 ], "packet_id": 34, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 951.2715096153848, "end": 953.3915096153848, "text": "但模型一旦开始生成内容", "raw_asr_text": "但模型一旦开始生成内容", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 34, "words": [ { "word": "但", "start": 951.2715096153848, "end": 951.5715096153848, "probability": 0.99755859375 }, { "word": "模", "start": 951.5715096153848, "end": 951.8115096153848, "probability": 0.9990234375 }, { "word": "型", "start": 951.8115096153848, "end": 952.0115096153849, "probability": 0.99951171875 }, { "word": "一", "start": 952.0115096153849, "end": 952.2315096153848, "probability": 0.9990234375 }, { "word": "旦", "start": 952.2315096153848, "end": 952.4315096153848, "probability": 0.9931640625 }, { "word": "开始", "start": 952.4315096153848, "end": 952.7515096153847, "probability": 0.9951171875 }, { "word": "生", "start": 952.7515096153847, "end": 952.8915096153848, "probability": 0.99267578125 }, { "word": "成", "start": 952.8915096153848, "end": 953.0315096153848, "probability": 0.98779296875 }, { "word": "内", "start": 953.0315096153848, "end": 953.1715096153848, "probability": 0.99755859375 }, { "word": "容", "start": 953.1715096153848, "end": 953.3915096153848, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10110379634291361, "compression_ratio": 1.2724719101123596, "no_speech_prob": 6.470405461422857e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "1仍然属于MOE模型显然它不会为每个Token都激活全部5520亿参数但借助CED与填充阶段的输入Token只需经过20层编码器因此DeepSeq在这阶段仅激活约80亿个参数但模型一旦开始生成内容每个新Token都必须同时经过编码器和解码器激活参数量随之升至160亿所以即便整个模型的总规模几乎是VS Flash的两倍模型中负责大部分繁重输入读取工作的部分其实非常难" } ], "id": 452, "review_annotation": "", "row_id": 452, "bounded_row_id": 452, "parent_row_ids": [ 452 ], "parent_row_id": 452, "packet_id": 34, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 937.1915096153848, "packet_end": 966.3794951923077, "raw_text": "但模型一旦开始生成内容", "selected_text": "但模型一旦开始生成内容", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 452 ], "packet_id": 34, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 953.3915096153848, "end": 957.1115096153848, "text": "每个新Token都必须同时经过编码器和解码器", "raw_asr_text": "每个新Token都必须同时经过编码器和解码器", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 34, "words": [ { "word": "每", "start": 953.3915096153848, "end": 953.7315096153848, "probability": 0.9697265625 }, { "word": "个", "start": 953.7315096153848, "end": 953.9115096153848, "probability": 0.99755859375 }, { "word": "新", "start": 953.9115096153848, "end": 954.1115096153848, "probability": 0.99609375 }, { "word": "T", "start": 954.1115096153848, "end": 954.2315096153848, "probability": 0.98828125 }, { "word": "oken", "start": 954.2315096153848, "end": 954.4315096153848, "probability": 0.99951171875 }, { "word": "都", "start": 954.4315096153848, "end": 954.7315096153848, "probability": 0.9873046875 }, { "word": "必", "start": 954.7315096153848, "end": 954.9715096153848, "probability": 0.9970703125 }, { "word": "须", "start": 954.9715096153848, "end": 955.1515096153848, "probability": 0.99951171875 }, { "word": "同", "start": 955.1515096153848, "end": 955.3315096153848, "probability": 0.99462890625 }, { "word": "时", "start": 955.3315096153848, "end": 955.6515096153848, "probability": 0.9990234375 }, { "word": "经", "start": 955.6515096153848, "end": 955.7915096153848, "probability": 0.9951171875 }, { "word": "过", "start": 955.7915096153848, "end": 955.9515096153848, "probability": 0.99853515625 }, { "word": "编", "start": 955.9515096153848, "end": 956.1115096153848, "probability": 0.99951171875 }, { "word": "码", "start": 956.1115096153848, "end": 956.2915096153848, "probability": 1.0 }, { "word": "器", "start": 956.2915096153848, "end": 956.4515096153848, "probability": 0.99951171875 }, { "word": "和", "start": 956.4515096153848, "end": 956.6315096153849, "probability": 0.99755859375 }, { "word": "解", "start": 956.6315096153849, "end": 956.7915096153848, "probability": 0.87890625 }, { "word": "码", "start": 956.7915096153848, "end": 956.9515096153848, "probability": 1.0 }, { "word": "器", "start": 956.9515096153848, "end": 957.1115096153848, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10110379634291361, "compression_ratio": 1.2724719101123596, "no_speech_prob": 6.470405461422857e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "1仍然属于MOE模型显然它不会为每个Token都激活全部5520亿参数但借助CED与填充阶段的输入Token只需经过20层编码器因此DeepSeq在这阶段仅激活约80亿个参数但模型一旦开始生成内容每个新Token都必须同时经过编码器和解码器激活参数量随之升至160亿所以即便整个模型的总规模几乎是VS Flash的两倍模型中负责大部分繁重输入读取工作的部分其实非常难" } ], "id": 453, "review_annotation": "", "row_id": 453, "bounded_row_id": 453, "parent_row_ids": [ 453 ], "parent_row_id": 453, "packet_id": 34, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 937.1915096153848, "packet_end": 966.3794951923077, "raw_text": "每个新Token都必须同时经过编码器和解码器", "selected_text": "每个新Token都必须同时经过编码器和解码器", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 453 ], "packet_id": 34, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 957.1115096153848, "end": 959.3115096153848, "text": "激活参数量随之升至160亿", "raw_asr_text": "激活参数量随之升至160亿", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 34, "words": [ { "word": "激", "start": 957.1115096153848, "end": 957.3715096153848, "probability": 0.998046875 }, { "word": "活", "start": 957.3715096153848, "end": 957.5115096153849, "probability": 1.0 }, { "word": "参", "start": 957.5115096153849, "end": 957.7115096153848, "probability": 0.998046875 }, { "word": "数", "start": 957.7115096153848, "end": 957.8515096153848, "probability": 1.0 }, { "word": "量", "start": 957.8515096153848, "end": 957.9915096153848, "probability": 0.9921875 }, { "word": "随", "start": 957.9915096153848, "end": 958.2315096153848, "probability": 0.998046875 }, { "word": "之", "start": 958.2315096153848, "end": 958.3515096153848, "probability": 0.94287109375 }, { "word": "升", "start": 958.3515096153848, "end": 958.4715096153848, "probability": 0.9814453125 }, { "word": "至", "start": 958.4715096153848, "end": 958.6715096153848, "probability": 0.99853515625 }, { "word": "160", "start": 958.6715096153848, "end": 959.0515096153848, "probability": 0.99609375 }, { "word": "亿", "start": 959.0515096153848, "end": 959.3115096153848, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10110379634291361, "compression_ratio": 1.2724719101123596, "no_speech_prob": 6.470405461422857e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "1仍然属于MOE模型显然它不会为每个Token都激活全部5520亿参数但借助CED与填充阶段的输入Token只需经过20层编码器因此DeepSeq在这阶段仅激活约80亿个参数但模型一旦开始生成内容每个新Token都必须同时经过编码器和解码器激活参数量随之升至160亿所以即便整个模型的总规模几乎是VS Flash的两倍模型中负责大部分繁重输入读取工作的部分其实非常难" } ], "id": 454, "review_annotation": "", "row_id": 454, "bounded_row_id": 454, "parent_row_ids": [ 454 ], "parent_row_id": 454, "packet_id": 34, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 937.1915096153848, "packet_end": 966.3794951923077, "raw_text": "激活参数量随之升至160亿", "selected_text": "激活参数量随之升至160亿", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 454 ], "packet_id": 34, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 959.3115096153848, "end": 962.6115096153848, "text": "所以即便整个模型的总规模几乎是VS Flash的两倍", "raw_asr_text": "所以即便整个模型的总规模几乎是VS Flash的两倍", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 34, "words": [ { "word": "所以", "start": 959.3115096153848, "end": 959.6115096153848, "probability": 0.99072265625 }, { "word": "即", "start": 959.6115096153848, "end": 959.7915096153848, "probability": 0.99755859375 }, { "word": "便", "start": 959.7915096153848, "end": 959.9315096153848, "probability": 0.998046875 }, { "word": "整", "start": 959.9315096153848, "end": 960.1315096153849, "probability": 0.9990234375 }, { "word": "个", "start": 960.1315096153849, "end": 960.2715096153848, "probability": 0.99951171875 }, { "word": "模", "start": 960.2715096153848, "end": 960.3715096153848, "probability": 0.99951171875 }, { "word": "型", "start": 960.3715096153848, "end": 960.4915096153848, "probability": 1.0 }, { "word": "的", "start": 960.4915096153848, "end": 960.6115096153848, "probability": 0.99853515625 }, { "word": "总", "start": 960.6115096153848, "end": 960.7315096153848, "probability": 0.9990234375 }, { "word": "规", "start": 960.7315096153848, "end": 960.8915096153848, "probability": 1.0 }, { "word": "模", "start": 960.8915096153848, "end": 961.0315096153848, "probability": 0.99951171875 }, { "word": "几", "start": 961.0315096153848, "end": 961.3315096153848, "probability": 0.99853515625 }, { "word": "乎", "start": 961.3315096153848, "end": 961.4115096153848, "probability": 1.0 }, { "word": "是", "start": 961.4115096153848, "end": 961.5315096153848, "probability": 0.99072265625 }, { "word": "V", "start": 961.5315096153848, "end": 961.6515096153848, "probability": 0.93310546875 }, { "word": "S", "start": 961.6515096153848, "end": 961.7915096153848, "probability": 0.6953125 }, { "word": " Flash", "start": 961.7915096153848, "end": 962.0515096153848, "probability": 0.6650390625 }, { "word": "的", "start": 962.0515096153848, "end": 962.2115096153848, "probability": 0.9931640625 }, { "word": "两", "start": 962.2115096153848, "end": 962.3715096153848, "probability": 0.70556640625 }, { "word": "倍", "start": 962.3715096153848, "end": 962.6115096153848, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10110379634291361, "compression_ratio": 1.2724719101123596, "no_speech_prob": 6.470405461422857e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "1仍然属于MOE模型显然它不会为每个Token都激活全部5520亿参数但借助CED与填充阶段的输入Token只需经过20层编码器因此DeepSeq在这阶段仅激活约80亿个参数但模型一旦开始生成内容每个新Token都必须同时经过编码器和解码器激活参数量随之升至160亿所以即便整个模型的总规模几乎是VS Flash的两倍模型中负责大部分繁重输入读取工作的部分其实非常难" } ], "id": 455, "review_annotation": "", "row_id": 455, "bounded_row_id": 455, "parent_row_ids": [ 455 ], "parent_row_id": 455, "packet_id": 34, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 937.1915096153848, "packet_end": 966.3794951923077, "raw_text": "所以即便整个模型的总规模几乎是VS Flash的两倍", "selected_text": "所以即便整个模型的总规模几乎是VS Flash的两倍", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 455 ], "packet_id": 34, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 962.6115096153848, "end": 964.8515096153848, "text": "模型中负责大部分繁重输入", "raw_asr_text": "模型中负责大部分繁重输入", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 34, "words": [ { "word": "模", "start": 962.6115096153848, "end": 963.0715096153848, "probability": 0.99072265625 }, { "word": "型", "start": 963.0715096153848, "end": 963.2315096153848, "probability": 1.0 }, { "word": "中", "start": 963.2315096153848, "end": 963.3915096153848, "probability": 0.99609375 }, { "word": "负", "start": 963.3915096153848, "end": 963.6115096153848, "probability": 0.9990234375 }, { "word": "责", "start": 963.6115096153848, "end": 963.7115096153848, "probability": 0.9990234375 }, { "word": "大", "start": 963.7115096153848, "end": 963.8715096153848, "probability": 0.998046875 }, { "word": "部分", "start": 963.8715096153848, "end": 964.1115096153848, "probability": 0.99853515625 }, { "word": "繁", "start": 964.1115096153848, "end": 964.3515096153848, "probability": 0.998046875 }, { "word": "重", "start": 964.3515096153848, "end": 964.4915096153848, "probability": 0.9814453125 }, { "word": "输", "start": 964.4915096153848, "end": 964.6915096153848, "probability": 0.9990234375 }, { "word": "入", "start": 964.6915096153848, "end": 964.8515096153848, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.10110379634291361, "compression_ratio": 1.2724719101123596, "no_speech_prob": 6.470405461422857e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "1仍然属于MOE模型显然它不会为每个Token都激活全部5520亿参数但借助CED与填充阶段的输入Token只需经过20层编码器因此DeepSeq在这阶段仅激活约80亿个参数但模型一旦开始生成内容每个新Token都必须同时经过编码器和解码器激活参数量随之升至160亿所以即便整个模型的总规模几乎是VS Flash的两倍模型中负责大部分繁重输入读取工作的部分其实非常难" } ], "id": 456, "review_annotation": "", "row_id": 456, "bounded_row_id": 456, "parent_row_ids": [ 456 ], "parent_row_id": 456, "packet_id": 34, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 937.1915096153848, "packet_end": 966.3794951923077, "raw_text": "模型中负责大部分繁重输入", "selected_text": "模型中负责大部分繁重输入", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 456 ], "packet_id": 34, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 965.5794951923077, "end": 966.6194951923078, "text": "读取工作的部分其实非常小", "raw_asr_text": "读取工作的部分其实非常小", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 35, "words": [ { "word": "读", "start": 964.7794951923078, "end": 965.0194951923078, "probability": 0.9658203125 }, { "word": "取", "start": 965.0194951923078, "end": 965.1794951923077, "probability": 0.9990234375 }, { "word": "工作", "start": 965.1794951923077, "end": 965.4394951923077, "probability": 0.9990234375 }, { "word": "的", "start": 965.4394951923077, "end": 965.5794951923077, "probability": 0.9970703125 }, { "word": "部分", "start": 965.5794951923077, "end": 965.7794951923078, "probability": 0.99853515625 }, { "word": "其实", "start": 965.7794951923078, "end": 966.0594951923077, "probability": 0.99658203125 }, { "word": "非常", "start": 966.0594951923077, "end": 966.3194951923077, "probability": 0.9990234375 }, { "word": "小", "start": 966.3194951923077, "end": 966.6194951923078, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09012821742466517, "compression_ratio": 1.2881844380403458, "no_speech_prob": 2.3532637644096965e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "读取工作的部分其实非常小这对智能体场景非常友好因为这里场景往往要读取几十万token最终却只生成几千个token的内容还有点不可思议的是哪怕上下文从4000token提升到100万token也就是上下文长度扩大到原来的256倍单token解码所需算力也仅从约20加全GFLOPs甚至25加全GFLOPs我认为这次版本更新的所有改动力上述这些新增核心组件是最值得关注的到了" } ], "id": 457, "review_annotation": "", "row_id": 457, "bounded_row_id": 457, "parent_row_ids": [ 457 ], "parent_row_id": 457, "packet_id": 35, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 964.7794951923078, "packet_end": 993.9674807692307, "raw_text": "读取工作的部分其实非常小", "selected_text": "读取工作的部分其实非常小", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 457 ], "packet_id": 35, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 966.6194951923078, "end": 968.5394951923078, "text": "这对智能体场景非常友好", "raw_asr_text": "这对智能体场景非常友好", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 35, "words": [ { "word": "这", "start": 966.6194951923078, "end": 966.9994951923078, "probability": 0.98876953125 }, { "word": "对", "start": 966.9994951923078, "end": 967.1194951923078, "probability": 1.0 }, { "word": "智", "start": 967.1194951923078, "end": 967.2794951923078, "probability": 0.998046875 }, { "word": "能", "start": 967.2794951923078, "end": 967.3794951923078, "probability": 0.998046875 }, { "word": "体", "start": 967.3794951923078, "end": 967.5394951923078, "probability": 0.9990234375 }, { "word": "场", "start": 967.5394951923078, "end": 967.7194951923078, "probability": 0.99951171875 }, { "word": "景", "start": 967.7194951923078, "end": 967.8594951923078, "probability": 1.0 }, { "word": "非常", "start": 967.8594951923078, "end": 968.1794951923077, "probability": 0.9990234375 }, { "word": "友", "start": 968.1794951923077, "end": 968.3594951923078, "probability": 0.9970703125 }, { "word": "好", "start": 968.3594951923078, "end": 968.5394951923078, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.09012821742466517, "compression_ratio": 1.2881844380403458, "no_speech_prob": 2.3532637644096965e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "读取工作的部分其实非常小这对智能体场景非常友好因为这里场景往往要读取几十万token最终却只生成几千个token的内容还有点不可思议的是哪怕上下文从4000token提升到100万token也就是上下文长度扩大到原来的256倍单token解码所需算力也仅从约20加全GFLOPs甚至25加全GFLOPs我认为这次版本更新的所有改动力上述这些新增核心组件是最值得关注的到了" } ], "id": 458, "review_annotation": "", "row_id": 458, "bounded_row_id": 458, "parent_row_ids": [ 458 ], "parent_row_id": 458, "packet_id": 35, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 964.7794951923078, "packet_end": 993.9674807692307, "raw_text": "这对智能体场景非常友好", "selected_text": "这对智能体场景非常友好", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 458 ], "packet_id": 35, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 968.5394951923078, "end": 971.2394951923078, "text": "因为这里场景往往要读取几十万token", "raw_asr_text": "因为这里场景往往要读取几十万token", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 35, "words": [ { "word": "因为", "start": 968.5394951923078, "end": 968.8394951923077, "probability": 0.99755859375 }, { "word": "这里", "start": 968.8394951923077, "end": 969.0594951923077, "probability": 0.66064453125 }, { "word": "场", "start": 969.0594951923077, "end": 969.2594951923078, "probability": 0.99267578125 }, { "word": "景", "start": 969.2594951923078, "end": 969.4394951923077, "probability": 1.0 }, { "word": "往", "start": 969.4394951923077, "end": 969.6394951923078, "probability": 0.99951171875 }, { "word": "往", "start": 969.6394951923078, "end": 969.7994951923077, "probability": 1.0 }, { "word": "要", "start": 969.7994951923077, "end": 969.9594951923077, "probability": 0.99951171875 }, { "word": "读", "start": 969.9594951923077, "end": 970.1794951923077, "probability": 0.9990234375 }, { "word": "取", "start": 970.1794951923077, "end": 970.3594951923078, "probability": 1.0 }, { "word": "几", "start": 970.3594951923078, "end": 970.5994951923078, "probability": 0.998046875 }, { "word": "十", "start": 970.5994951923078, "end": 970.7194951923078, "probability": 0.99609375 }, { "word": "万", "start": 970.7194951923078, "end": 970.8794951923078, "probability": 0.99951171875 }, { "word": "t", "start": 970.8794951923078, "end": 971.0394951923078, "probability": 0.5498046875 }, { "word": "oken", "start": 971.0394951923078, "end": 971.2394951923078, "probability": 0.54248046875 } ], "asr_metrics": { "avg_logprob": -0.09012821742466517, "compression_ratio": 1.2881844380403458, "no_speech_prob": 2.3532637644096965e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "读取工作的部分其实非常小这对智能体场景非常友好因为这里场景往往要读取几十万token最终却只生成几千个token的内容还有点不可思议的是哪怕上下文从4000token提升到100万token也就是上下文长度扩大到原来的256倍单token解码所需算力也仅从约20加全GFLOPs甚至25加全GFLOPs我认为这次版本更新的所有改动力上述这些新增核心组件是最值得关注的到了" } ], "id": 459, "review_annotation": "", "row_id": 459, "bounded_row_id": 459, "parent_row_ids": [ 459 ], "parent_row_id": 459, "packet_id": 35, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 964.7794951923078, "packet_end": 993.9674807692307, "raw_text": "因为这里场景往往要读取几十万token", "selected_text": "因为这里场景往往要读取几十万token", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 459 ], "packet_id": 35, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 971.2394951923078, "end": 973.8394951923077, "text": "最终却只生成几千个token的内容", "raw_asr_text": "最终却只生成几千个token的内容", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 35, "words": [ { "word": "最", "start": 971.2394951923078, "end": 971.5594951923077, "probability": 0.96875 }, { "word": "终", "start": 971.5594951923077, "end": 971.7594951923078, "probability": 0.9990234375 }, { "word": "却", "start": 971.7594951923078, "end": 972.0194951923078, "probability": 0.99951171875 }, { "word": "只", "start": 972.0194951923078, "end": 972.1794951923077, "probability": 0.99951171875 }, { "word": "生", "start": 972.1794951923077, "end": 972.3394951923077, "probability": 0.986328125 }, { "word": "成", "start": 972.3394951923077, "end": 972.4994951923078, "probability": 0.99853515625 }, { "word": "几", "start": 972.4994951923078, "end": 972.7994951923077, "probability": 0.9990234375 }, { "word": "千", "start": 972.7994951923077, "end": 972.9994951923078, "probability": 0.9990234375 }, { "word": "个", "start": 972.9994951923078, "end": 973.1394951923078, "probability": 0.9990234375 }, { "word": "t", "start": 973.1394951923078, "end": 973.2794951923078, "probability": 0.98828125 }, { "word": "oken", "start": 973.2794951923078, "end": 973.4194951923077, "probability": 0.99951171875 }, { "word": "的", "start": 973.4194951923077, "end": 973.5394951923078, "probability": 0.998046875 }, { "word": "内", "start": 973.5394951923078, "end": 973.6794951923077, "probability": 0.99951171875 }, { "word": "容", "start": 973.6794951923077, "end": 973.8394951923077, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09012821742466517, "compression_ratio": 1.2881844380403458, "no_speech_prob": 2.3532637644096965e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "读取工作的部分其实非常小这对智能体场景非常友好因为这里场景往往要读取几十万token最终却只生成几千个token的内容还有点不可思议的是哪怕上下文从4000token提升到100万token也就是上下文长度扩大到原来的256倍单token解码所需算力也仅从约20加全GFLOPs甚至25加全GFLOPs我认为这次版本更新的所有改动力上述这些新增核心组件是最值得关注的到了" } ], "id": 460, "review_annotation": "", "row_id": 460, "bounded_row_id": 460, "parent_row_ids": [ 460 ], "parent_row_id": 460, "packet_id": 35, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 964.7794951923078, "packet_end": 993.9674807692307, "raw_text": "最终却只生成几千个token的内容", "selected_text": "最终却只生成几千个token的内容", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 460 ], "packet_id": 35, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 973.8394951923077, "end": 975.2994951923077, "text": "还有点不可思议的是", "raw_asr_text": "还有点不可思议的是", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 35, "words": [ { "word": "还有", "start": 973.8394951923077, "end": 974.1794951923077, "probability": 0.99169921875 }, { "word": "点", "start": 974.1794951923077, "end": 974.4194951923077, "probability": 0.99951171875 }, { "word": "不", "start": 974.4194951923077, "end": 974.5994951923078, "probability": 1.0 }, { "word": "可", "start": 974.5994951923078, "end": 974.7594951923078, "probability": 0.99951171875 }, { "word": "思", "start": 974.7594951923078, "end": 974.9194951923077, "probability": 0.99951171875 }, { "word": "议", "start": 974.9194951923077, "end": 975.0394951923078, "probability": 0.998046875 }, { "word": "的是", "start": 975.0394951923078, "end": 975.2994951923077, "probability": 0.99462890625 } ], "asr_metrics": { "avg_logprob": -0.09012821742466517, "compression_ratio": 1.2881844380403458, "no_speech_prob": 2.3532637644096965e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "读取工作的部分其实非常小这对智能体场景非常友好因为这里场景往往要读取几十万token最终却只生成几千个token的内容还有点不可思议的是哪怕上下文从4000token提升到100万token也就是上下文长度扩大到原来的256倍单token解码所需算力也仅从约20加全GFLOPs甚至25加全GFLOPs我认为这次版本更新的所有改动力上述这些新增核心组件是最值得关注的到了" } ], "id": 461, "review_annotation": "", "row_id": 461, "bounded_row_id": 461, "parent_row_ids": [ 461 ], "parent_row_id": 461, "packet_id": 35, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 964.7794951923078, "packet_end": 993.9674807692307, "raw_text": "还有点不可思议的是", "selected_text": "还有点不可思议的是", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 461 ], "packet_id": 35, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 975.2994951923077, "end": 978.9594951923077, "text": "哪怕上下文从4000token提升到100万token", "raw_asr_text": "哪怕上下文从4000token提升到100万token", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 35, "words": [ { "word": "哪", "start": 975.2994951923077, "end": 975.6594951923078, "probability": 0.9970703125 }, { "word": "怕", "start": 975.6594951923078, "end": 975.8394951923077, "probability": 1.0 }, { "word": "上", "start": 975.8394951923077, "end": 976.0194951923078, "probability": 0.99951171875 }, { "word": "下", "start": 976.0194951923078, "end": 976.1994951923077, "probability": 0.99951171875 }, { "word": "文", "start": 976.1994951923077, "end": 976.3394951923077, "probability": 0.9462890625 }, { "word": "从", "start": 976.3394951923077, "end": 976.5394951923078, "probability": 0.9990234375 }, { "word": "4", "start": 976.5394951923078, "end": 976.7994951923077, "probability": 0.97265625 }, { "word": "000", "start": 976.7994951923077, "end": 976.9794951923078, "probability": 0.935546875 }, { "word": "t", "start": 976.9794951923078, "end": 977.1194951923078, "probability": 0.98193359375 }, { "word": "oken", "start": 977.1194951923078, "end": 977.2994951923077, "probability": 0.80322265625 }, { "word": "提", "start": 977.2994951923077, "end": 977.5794951923077, "probability": 0.99658203125 }, { "word": "升", "start": 977.5794951923077, "end": 977.7594951923078, "probability": 0.998046875 }, { "word": "到", "start": 977.7594951923078, "end": 977.8994951923078, "probability": 0.9990234375 }, { "word": "100", "start": 977.8994951923078, "end": 978.2794951923078, "probability": 0.9951171875 }, { "word": "万", "start": 978.2794951923078, "end": 978.5394951923078, "probability": 0.9990234375 }, { "word": "t", "start": 978.5394951923078, "end": 978.7594951923078, "probability": 0.998046875 }, { "word": "oken", "start": 978.7594951923078, "end": 978.9594951923077, "probability": 0.98388671875 } ], "asr_metrics": { "avg_logprob": -0.09012821742466517, "compression_ratio": 1.2881844380403458, "no_speech_prob": 2.3532637644096965e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "读取工作的部分其实非常小这对智能体场景非常友好因为这里场景往往要读取几十万token最终却只生成几千个token的内容还有点不可思议的是哪怕上下文从4000token提升到100万token也就是上下文长度扩大到原来的256倍单token解码所需算力也仅从约20加全GFLOPs甚至25加全GFLOPs我认为这次版本更新的所有改动力上述这些新增核心组件是最值得关注的到了" } ], "id": 462, "review_annotation": "", "row_id": 462, "bounded_row_id": 462, "parent_row_ids": [ 462 ], "parent_row_id": 462, "packet_id": 35, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 964.7794951923078, "packet_end": 993.9674807692307, "raw_text": "哪怕上下文从4000token提升到100万token", "selected_text": "哪怕上下文从4000token提升到100万token", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 462 ], "packet_id": 35, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 978.9594951923077, "end": 982.3394951923077, "text": "也就是上下文长度扩大到原来的256倍", "raw_asr_text": "也就是上下文长度扩大到原来的256倍", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 35, "words": [ { "word": "也", "start": 978.9594951923077, "end": 979.2994951923077, "probability": 0.9970703125 }, { "word": "就是", "start": 979.2994951923077, "end": 979.5194951923078, "probability": 0.99951171875 }, { "word": "上", "start": 979.5194951923078, "end": 979.7794951923078, "probability": 0.99951171875 }, { "word": "下", "start": 979.7794951923078, "end": 979.9594951923077, "probability": 1.0 }, { "word": "文", "start": 979.9594951923077, "end": 980.1394951923078, "probability": 0.99951171875 }, { "word": "长", "start": 980.1394951923078, "end": 980.3394951923077, "probability": 0.99755859375 }, { "word": "度", "start": 980.3394951923077, "end": 980.5994951923078, "probability": 1.0 }, { "word": "扩", "start": 980.5994951923078, "end": 980.7994951923077, "probability": 0.99853515625 }, { "word": "大", "start": 980.7994951923077, "end": 980.9394951923077, "probability": 0.9990234375 }, { "word": "到", "start": 980.9394951923077, "end": 981.0994951923078, "probability": 0.99951171875 }, { "word": "原", "start": 981.0994951923078, "end": 981.2994951923077, "probability": 0.9990234375 }, { "word": "来", "start": 981.2994951923077, "end": 981.5194951923078, "probability": 1.0 }, { "word": "的", "start": 981.5194951923078, "end": 981.6394951923078, "probability": 1.0 }, { "word": "25", "start": 981.6394951923078, "end": 981.8794951923078, "probability": 0.986328125 }, { "word": "6", "start": 981.8794951923078, "end": 982.1394951923078, "probability": 0.9912109375 }, { "word": "倍", "start": 982.1394951923078, "end": 982.3394951923077, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09012821742466517, "compression_ratio": 1.2881844380403458, "no_speech_prob": 2.3532637644096965e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "读取工作的部分其实非常小这对智能体场景非常友好因为这里场景往往要读取几十万token最终却只生成几千个token的内容还有点不可思议的是哪怕上下文从4000token提升到100万token也就是上下文长度扩大到原来的256倍单token解码所需算力也仅从约20加全GFLOPs甚至25加全GFLOPs我认为这次版本更新的所有改动力上述这些新增核心组件是最值得关注的到了" } ], "id": 463, "review_annotation": "", "row_id": 463, "bounded_row_id": 463, "parent_row_ids": [ 463 ], "parent_row_id": 463, "packet_id": 35, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 964.7794951923078, "packet_end": 993.9674807692307, "raw_text": "也就是上下文长度扩大到原来的256倍", "selected_text": "也就是上下文长度扩大到原来的256倍", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 463 ], "packet_id": 35, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 982.3394951923077, "end": 984.1394951923078, "text": "单token解码所需算力", "raw_asr_text": "单token解码所需算力", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 35, "words": [ { "word": "单", "start": 982.3394951923077, "end": 982.8194951923077, "probability": 0.9892578125 }, { "word": "t", "start": 982.8194951923077, "end": 982.9594951923077, "probability": 0.97998046875 }, { "word": "oken", "start": 982.9594951923077, "end": 983.1194951923078, "probability": 0.9990234375 }, { "word": "解", "start": 983.1194951923078, "end": 983.2594951923078, "probability": 0.87451171875 }, { "word": "码", "start": 983.2594951923078, "end": 983.4794951923078, "probability": 0.9970703125 }, { "word": "所", "start": 983.4794951923078, "end": 983.6794951923077, "probability": 0.99560546875 }, { "word": "需", "start": 983.6794951923077, "end": 983.8194951923077, "probability": 0.9990234375 }, { "word": "算", "start": 983.8194951923077, "end": 983.9794951923078, "probability": 0.99951171875 }, { "word": "力", "start": 983.9794951923078, "end": 984.1394951923078, "probability": 0.97900390625 } ], "asr_metrics": { "avg_logprob": -0.09012821742466517, "compression_ratio": 1.2881844380403458, "no_speech_prob": 2.3532637644096965e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "读取工作的部分其实非常小这对智能体场景非常友好因为这里场景往往要读取几十万token最终却只生成几千个token的内容还有点不可思议的是哪怕上下文从4000token提升到100万token也就是上下文长度扩大到原来的256倍单token解码所需算力也仅从约20加全GFLOPs甚至25加全GFLOPs我认为这次版本更新的所有改动力上述这些新增核心组件是最值得关注的到了" } ], "id": 464, "review_annotation": "", "row_id": 464, "bounded_row_id": 464, "parent_row_ids": [ 464 ], "parent_row_id": 464, "packet_id": 35, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 964.7794951923078, "packet_end": 993.9674807692307, "raw_text": "单token解码所需算力", "selected_text": "单token解码所需算力", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 464 ], "packet_id": 35, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 984.1394951923078, "end": 986.2394951923078, "text": "也仅从约20加全GFLOPs", "raw_asr_text": "也仅从约20加全GFLOPs", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 35, "words": [ { "word": "也", "start": 984.1394951923078, "end": 984.4394951923077, "probability": 0.998046875 }, { "word": "仅", "start": 984.4394951923077, "end": 984.6394951923078, "probability": 0.9970703125 }, { "word": "从", "start": 984.6394951923078, "end": 984.7994951923077, "probability": 0.99853515625 }, { "word": "约", "start": 984.7994951923077, "end": 984.9994951923078, "probability": 0.99609375 }, { "word": "20", "start": 984.9994951923078, "end": 985.3794951923078, "probability": 0.9970703125 }, { "word": "加", "start": 985.3794951923078, "end": 985.5594951923077, "probability": 0.6943359375 }, { "word": "全", "start": 985.5594951923077, "end": 985.7194951923078, "probability": 0.7236328125 }, { "word": "G", "start": 985.7194951923078, "end": 985.8394951923077, "probability": 0.8671875 }, { "word": "FL", "start": 985.8394951923077, "end": 985.9794951923078, "probability": 0.66748046875 }, { "word": "OP", "start": 985.9794951923078, "end": 986.1194951923078, "probability": 0.91064453125 }, { "word": "s", "start": 986.1194951923078, "end": 986.2394951923078, "probability": 0.54052734375 } ], "asr_metrics": { "avg_logprob": -0.09012821742466517, "compression_ratio": 1.2881844380403458, "no_speech_prob": 2.3532637644096965e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "读取工作的部分其实非常小这对智能体场景非常友好因为这里场景往往要读取几十万token最终却只生成几千个token的内容还有点不可思议的是哪怕上下文从4000token提升到100万token也就是上下文长度扩大到原来的256倍单token解码所需算力也仅从约20加全GFLOPs甚至25加全GFLOPs我认为这次版本更新的所有改动力上述这些新增核心组件是最值得关注的到了" } ], "id": 465, "review_annotation": "", "row_id": 465, "bounded_row_id": 465, "parent_row_ids": [ 465 ], "parent_row_id": 465, "packet_id": 35, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 964.7794951923078, "packet_end": 993.9674807692307, "raw_text": "也仅从约20加全GFLOPs", "selected_text": "也仅从约20加全GFLOPs", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 465 ], "packet_id": 35, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 986.2394951923078, "end": 987.9994951923078, "text": "甚至25加全GFLOPs", "raw_asr_text": "甚至25加全GFLOPs", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 35, "words": [ { "word": "甚", "start": 986.2394951923078, "end": 986.4594951923077, "probability": 0.5322265625 }, { "word": "至", "start": 986.4594951923077, "end": 986.6594951923078, "probability": 1.0 }, { "word": "25", "start": 986.6594951923078, "end": 987.0794951923077, "probability": 0.9990234375 }, { "word": "加", "start": 987.0794951923077, "end": 987.3594951923078, "probability": 0.9990234375 }, { "word": "全", "start": 987.3594951923078, "end": 987.4594951923077, "probability": 0.998046875 }, { "word": "G", "start": 987.4594951923077, "end": 987.5594951923077, "probability": 0.9990234375 }, { "word": "FL", "start": 987.5594951923077, "end": 987.7794951923078, "probability": 0.99853515625 }, { "word": "OP", "start": 987.7794951923078, "end": 987.8594951923078, "probability": 1.0 }, { "word": "s", "start": 987.8594951923078, "end": 987.9994951923078, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.09012821742466517, "compression_ratio": 1.2881844380403458, "no_speech_prob": 2.3532637644096965e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "读取工作的部分其实非常小这对智能体场景非常友好因为这里场景往往要读取几十万token最终却只生成几千个token的内容还有点不可思议的是哪怕上下文从4000token提升到100万token也就是上下文长度扩大到原来的256倍单token解码所需算力也仅从约20加全GFLOPs甚至25加全GFLOPs我认为这次版本更新的所有改动力上述这些新增核心组件是最值得关注的到了" } ], "id": 466, "review_annotation": "", "row_id": 466, "bounded_row_id": 466, "parent_row_ids": [ 466 ], "parent_row_id": 466, "packet_id": 35, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 964.7794951923078, "packet_end": 993.9674807692307, "raw_text": "甚至25加全GFLOPs", "selected_text": "甚至25加全GFLOPs", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 466 ], "packet_id": 35, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 987.9994951923078, "end": 990.3194951923077, "text": "我认为这次版本更新的所有改动力", "raw_asr_text": "我认为这次版本更新的所有改动力", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 35, "words": [ { "word": "我", "start": 987.9994951923078, "end": 988.2394951923078, "probability": 0.99853515625 }, { "word": "认", "start": 988.2394951923078, "end": 988.3794951923078, "probability": 1.0 }, { "word": "为", "start": 988.3794951923078, "end": 988.4994951923078, "probability": 1.0 }, { "word": "这", "start": 988.4994951923078, "end": 988.6794951923077, "probability": 0.99951171875 }, { "word": "次", "start": 988.6794951923077, "end": 988.7994951923077, "probability": 0.99951171875 }, { "word": "版", "start": 988.7994951923077, "end": 988.9594951923077, "probability": 0.9990234375 }, { "word": "本", "start": 988.9594951923077, "end": 989.1394951923078, "probability": 0.99951171875 }, { "word": "更", "start": 989.1394951923078, "end": 989.3194951923077, "probability": 0.99853515625 }, { "word": "新", "start": 989.3194951923077, "end": 989.4794951923078, "probability": 1.0 }, { "word": "的", "start": 989.4794951923078, "end": 989.5794951923077, "probability": 1.0 }, { "word": "所有", "start": 989.5794951923077, "end": 989.7994951923077, "probability": 0.99951171875 }, { "word": "改", "start": 989.7994951923077, "end": 989.9994951923078, "probability": 0.99853515625 }, { "word": "动", "start": 989.9994951923078, "end": 990.1794951923077, "probability": 0.99755859375 }, { "word": "力", "start": 990.1794951923077, "end": 990.3194951923077, "probability": 0.1768798828125 } ], "asr_metrics": { "avg_logprob": -0.09012821742466517, "compression_ratio": 1.2881844380403458, "no_speech_prob": 2.3532637644096965e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "读取工作的部分其实非常小这对智能体场景非常友好因为这里场景往往要读取几十万token最终却只生成几千个token的内容还有点不可思议的是哪怕上下文从4000token提升到100万token也就是上下文长度扩大到原来的256倍单token解码所需算力也仅从约20加全GFLOPs甚至25加全GFLOPs我认为这次版本更新的所有改动力上述这些新增核心组件是最值得关注的到了" } ], "id": 467, "review_annotation": "", "row_id": 467, "bounded_row_id": 467, "parent_row_ids": [ 467 ], "parent_row_id": 467, "packet_id": 35, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 964.7794951923078, "packet_end": 993.9674807692307, "raw_text": "我认为这次版本更新的所有改动力", "selected_text": "我认为这次版本更新的所有改动力", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 467 ], "packet_id": 35, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 990.3194951923077, "end": 993.1674807692308, "text": "上述这些新增核心组件是最值得关注的", "raw_asr_text": "上述这些新增核心组件是最值得关注的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 35, "words": [ { "word": "上", "start": 990.3194951923077, "end": 990.6194951923078, "probability": 0.98583984375 }, { "word": "述", "start": 990.6194951923078, "end": 990.7594951923078, "probability": 0.9990234375 }, { "word": "这", "start": 990.7594951923078, "end": 990.8994951923078, "probability": 0.99951171875 }, { "word": "些", "start": 990.8994951923078, "end": 991.0794951923077, "probability": 1.0 }, { "word": "新", "start": 991.0794951923077, "end": 991.3194951923077, "probability": 0.99951171875 }, { "word": "增", "start": 991.3194951923077, "end": 991.5194951923078, "probability": 1.0 }, { "word": "核", "start": 991.5194951923078, "end": 991.6994951923077, "probability": 0.99951171875 }, { "word": "心", "start": 991.6994951923077, "end": 991.8394951923077, "probability": 0.9990234375 }, { "word": "组", "start": 991.8394951923077, "end": 992.0194951923078, "probability": 0.9970703125 }, { "word": "件", "start": 992.0194951923078, "end": 992.1594951923078, "probability": 0.98046875 }, { "word": "是", "start": 992.1594951923078, "end": 992.4194951923077, "probability": 0.9990234375 }, { "word": "最", "start": 992.4194951923077, "end": 992.5994951923078, "probability": 0.9990234375 }, { "word": "值", "start": 992.5994951923078, "end": 992.7394951923078, "probability": 1.0 }, { "word": "得", "start": 992.7394951923078, "end": 992.8594951923078, "probability": 0.99267578125 }, { "word": "关", "start": 992.8594951923078, "end": 992.9994951923078, "probability": 0.9990234375 }, { "word": "注", "start": 992.9994951923078, "end": 993.1394951923078, "probability": 1.0 }, { "word": "的", "start": 993.1394951923078, "end": 993.2994951923077, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09012821742466517, "compression_ratio": 1.2881844380403458, "no_speech_prob": 2.3532637644096965e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "读取工作的部分其实非常小这对智能体场景非常友好因为这里场景往往要读取几十万token最终却只生成几千个token的内容还有点不可思议的是哪怕上下文从4000token提升到100万token也就是上下文长度扩大到原来的256倍单token解码所需算力也仅从约20加全GFLOPs甚至25加全GFLOPs我认为这次版本更新的所有改动力上述这些新增核心组件是最值得关注的到了" } ], "id": 468, "review_annotation": "", "row_id": 468, "bounded_row_id": 468, "parent_row_ids": [ 468 ], "parent_row_id": 468, "packet_id": 35, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 964.7794951923078, "packet_end": 993.9674807692307, "raw_text": "上述这些新增核心组件是最值得关注的", "selected_text": "上述这些新增核心组件是最值得关注的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 468 ], "packet_id": 35, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 993.2674807692308, "end": 994.4674807692309, "text": "到了这个阶段", "raw_asr_text": "到了这个阶段", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 36, "words": [ { "word": "到了", "start": 993.2674807692308, "end": 993.7874807692308, "probability": 0.92626953125 }, { "word": "这个", "start": 993.7874807692308, "end": 994.1074807692308, "probability": 0.9892578125 }, { "word": "阶", "start": 994.1074807692308, "end": 994.3074807692309, "probability": 0.99609375 }, { "word": "段", "start": 994.3074807692309, "end": 994.4674807692309, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.11299644470214844, "compression_ratio": 1.226775956284153, "no_speech_prob": 3.439483073353067e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "是最值得关注的到了这个阶段百万token上下文几乎已经不是这个模型最有冲击力的特性了因此Deepseek开始将注意力转向架构中其他人在浪费内存带宽和算力的环节接下来要讲的是Deepseek在V4.1中重新实现时同样做了优化的其他组件首先是MHC我之前已经做过MHC的相关讲解视频这里就长话短说V4.1针对它的专属改动叫做单路径MHC这种设计基本解决了" } ], "id": 469, "review_annotation": "", "row_id": 469, "bounded_row_id": 469, "parent_row_ids": [ 469 ], "parent_row_id": 469, "packet_id": 36, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 992.3674807692308, "packet_end": 1021.5554663461539, "raw_text": "到了这个阶段", "selected_text": "到了这个阶段", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 469 ], "packet_id": 36, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 994.4674807692309, "end": 996.9874807692308, "text": "百万token上下文几乎已经不是这个模型", "raw_asr_text": "百万token上下文几乎已经不是这个模型", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 36, "words": [ { "word": "百", "start": 994.4674807692309, "end": 994.7274807692309, "probability": 0.8837890625 }, { "word": "万", "start": 994.7274807692309, "end": 994.8874807692308, "probability": 0.9990234375 }, { "word": "t", "start": 994.8874807692308, "end": 995.0274807692308, "probability": 0.40625 }, { "word": "oken", "start": 995.0274807692308, "end": 995.1674807692308, "probability": 0.99462890625 }, { "word": "上", "start": 995.1674807692308, "end": 995.3274807692309, "probability": 0.978515625 }, { "word": "下", "start": 995.3274807692309, "end": 995.4674807692309, "probability": 0.99853515625 }, { "word": "文", "start": 995.4674807692309, "end": 995.6274807692308, "probability": 0.96875 }, { "word": "几", "start": 995.6274807692308, "end": 995.8674807692308, "probability": 0.9912109375 }, { "word": "乎", "start": 995.8674807692308, "end": 996.0274807692308, "probability": 1.0 }, { "word": "已经", "start": 996.0274807692308, "end": 996.2074807692309, "probability": 0.98681640625 }, { "word": "不是", "start": 996.2074807692309, "end": 996.5074807692308, "probability": 0.98681640625 }, { "word": "这个", "start": 996.5074807692308, "end": 996.7074807692309, "probability": 0.98486328125 }, { "word": "模", "start": 996.7074807692309, "end": 996.8474807692309, "probability": 0.96728515625 }, { "word": "型", "start": 996.8474807692309, "end": 996.9874807692308, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11299644470214844, "compression_ratio": 1.226775956284153, "no_speech_prob": 3.439483073353067e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "是最值得关注的到了这个阶段百万token上下文几乎已经不是这个模型最有冲击力的特性了因此Deepseek开始将注意力转向架构中其他人在浪费内存带宽和算力的环节接下来要讲的是Deepseek在V4.1中重新实现时同样做了优化的其他组件首先是MHC我之前已经做过MHC的相关讲解视频这里就长话短说V4.1针对它的专属改动叫做单路径MHC这种设计基本解决了" } ], "id": 470, "review_annotation": "", "row_id": 470, "bounded_row_id": 470, "parent_row_ids": [ 470 ], "parent_row_id": 470, "packet_id": 36, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 992.3674807692308, "packet_end": 1021.5554663461539, "raw_text": "百万token上下文几乎已经不是这个模型", "selected_text": "百万token上下文几乎已经不是这个模型", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 470 ], "packet_id": 36, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 996.9874807692308, "end": 998.2274807692309, "text": "最有冲击力的特性了", "raw_asr_text": "最有冲击力的特性了", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 36, "words": [ { "word": "最", "start": 996.9874807692308, "end": 997.1474807692308, "probability": 0.9853515625 }, { "word": "有", "start": 997.1474807692308, "end": 997.2674807692308, "probability": 0.99853515625 }, { "word": "冲", "start": 997.2674807692308, "end": 997.4274807692308, "probability": 0.99267578125 }, { "word": "击", "start": 997.4274807692308, "end": 997.5874807692309, "probability": 1.0 }, { "word": "力", "start": 997.5874807692309, "end": 997.6874807692309, "probability": 0.98779296875 }, { "word": "的", "start": 997.6874807692309, "end": 997.7874807692308, "probability": 0.99755859375 }, { "word": "特", "start": 997.7874807692308, "end": 997.9274807692308, "probability": 0.99755859375 }, { "word": "性", "start": 997.9274807692308, "end": 998.1074807692308, "probability": 0.9970703125 }, { "word": "了", "start": 998.1074807692308, "end": 998.2274807692309, "probability": 0.96826171875 } ], "asr_metrics": { "avg_logprob": -0.11299644470214844, "compression_ratio": 1.226775956284153, "no_speech_prob": 3.439483073353067e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "是最值得关注的到了这个阶段百万token上下文几乎已经不是这个模型最有冲击力的特性了因此Deepseek开始将注意力转向架构中其他人在浪费内存带宽和算力的环节接下来要讲的是Deepseek在V4.1中重新实现时同样做了优化的其他组件首先是MHC我之前已经做过MHC的相关讲解视频这里就长话短说V4.1针对它的专属改动叫做单路径MHC这种设计基本解决了" } ], "id": 471, "review_annotation": "", "row_id": 471, "bounded_row_id": 471, "parent_row_ids": [ 471 ], "parent_row_id": 471, "packet_id": 36, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 992.3674807692308, "packet_end": 1021.5554663461539, "raw_text": "最有冲击力的特性了", "selected_text": "最有冲击力的特性了", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 471 ], "packet_id": 36, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 998.2274807692309, "end": 1001.1074807692308, "text": "因此Deepseek开始将注意力转向架构中", "raw_asr_text": "因此Deepseek开始将注意力转向架构中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 36, "words": [ { "word": "因", "start": 998.2274807692309, "end": 998.4474807692309, "probability": 0.9951171875 }, { "word": "此", "start": 998.4474807692309, "end": 998.6474807692308, "probability": 1.0 }, { "word": "De", "start": 998.6474807692308, "end": 998.8274807692309, "probability": 0.458251953125 }, { "word": "ep", "start": 998.8274807692309, "end": 998.9074807692308, "probability": 0.94580078125 }, { "word": "se", "start": 998.9074807692308, "end": 999.0274807692308, "probability": 0.32861328125 }, { "word": "ek", "start": 999.0274807692308, "end": 999.1474807692308, "probability": 0.6533203125 }, { "word": "开始", "start": 999.1474807692308, "end": 999.4874807692308, "probability": 0.98486328125 }, { "word": "将", "start": 999.4874807692308, "end": 999.6874807692309, "probability": 0.9970703125 }, { "word": "注意", "start": 999.6874807692309, "end": 999.9074807692308, "probability": 0.99560546875 }, { "word": "力", "start": 999.9074807692308, "end": 1000.0674807692309, "probability": 0.99853515625 }, { "word": "转", "start": 1000.0674807692309, "end": 1000.3274807692309, "probability": 0.998046875 }, { "word": "向", "start": 1000.3274807692309, "end": 1000.5674807692309, "probability": 0.99951171875 }, { "word": "架", "start": 1000.5674807692309, "end": 1000.8074807692309, "probability": 0.9970703125 }, { "word": "构", "start": 1000.8074807692309, "end": 1000.9474807692309, "probability": 0.9990234375 }, { "word": "中", "start": 1000.9474807692309, "end": 1001.1074807692308, "probability": 0.990234375 } ], "asr_metrics": { "avg_logprob": -0.11299644470214844, "compression_ratio": 1.226775956284153, "no_speech_prob": 3.439483073353067e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "是最值得关注的到了这个阶段百万token上下文几乎已经不是这个模型最有冲击力的特性了因此Deepseek开始将注意力转向架构中其他人在浪费内存带宽和算力的环节接下来要讲的是Deepseek在V4.1中重新实现时同样做了优化的其他组件首先是MHC我之前已经做过MHC的相关讲解视频这里就长话短说V4.1针对它的专属改动叫做单路径MHC这种设计基本解决了" } ], "id": 472, "review_annotation": "", "row_id": 472, "bounded_row_id": 472, "parent_row_ids": [ 472 ], "parent_row_id": 472, "packet_id": 36, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 992.3674807692308, "packet_end": 1021.5554663461539, "raw_text": "因此Deepseek开始将注意力转向架构中", "selected_text": "因此Deepseek开始将注意力转向架构中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 472 ], "packet_id": 36, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1001.1074807692308, "end": 1004.3074807692309, "text": "其他人在浪费内存带宽和算力的环节", "raw_asr_text": "其他人在浪费内存带宽和算力的环节", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 36, "words": [ { "word": "其", "start": 1001.1074807692308, "end": 1001.4674807692309, "probability": 0.99365234375 }, { "word": "他", "start": 1001.4674807692309, "end": 1001.6674807692308, "probability": 0.95947265625 }, { "word": "人", "start": 1001.6674807692308, "end": 1001.8274807692309, "probability": 0.69384765625 }, { "word": "在", "start": 1001.8274807692309, "end": 1002.0274807692308, "probability": 0.9873046875 }, { "word": "浪", "start": 1002.0274807692308, "end": 1002.2874807692308, "probability": 0.998046875 }, { "word": "费", "start": 1002.2874807692308, "end": 1002.4674807692309, "probability": 1.0 }, { "word": "内", "start": 1002.4674807692309, "end": 1002.6074807692308, "probability": 0.9970703125 }, { "word": "存", "start": 1002.6074807692308, "end": 1002.8474807692309, "probability": 1.0 }, { "word": "带", "start": 1002.8474807692309, "end": 1003.1074807692308, "probability": 0.88623046875 }, { "word": "宽", "start": 1003.1074807692308, "end": 1003.3074807692309, "probability": 0.9990234375 }, { "word": "和", "start": 1003.3074807692309, "end": 1003.5074807692308, "probability": 0.99609375 }, { "word": "算", "start": 1003.5074807692308, "end": 1003.6874807692309, "probability": 0.99951171875 }, { "word": "力", "start": 1003.6874807692309, "end": 1003.8474807692309, "probability": 0.99267578125 }, { "word": "的", "start": 1003.8474807692309, "end": 1003.9474807692309, "probability": 0.99609375 }, { "word": "环", "start": 1003.9474807692309, "end": 1004.1274807692308, "probability": 0.9990234375 }, { "word": "节", "start": 1004.1274807692308, "end": 1004.3074807692309, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11299644470214844, "compression_ratio": 1.226775956284153, "no_speech_prob": 3.439483073353067e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "是最值得关注的到了这个阶段百万token上下文几乎已经不是这个模型最有冲击力的特性了因此Deepseek开始将注意力转向架构中其他人在浪费内存带宽和算力的环节接下来要讲的是Deepseek在V4.1中重新实现时同样做了优化的其他组件首先是MHC我之前已经做过MHC的相关讲解视频这里就长话短说V4.1针对它的专属改动叫做单路径MHC这种设计基本解决了" } ], "id": 473, "review_annotation": "", "row_id": 473, "bounded_row_id": 473, "parent_row_ids": [ 473 ], "parent_row_id": 473, "packet_id": 36, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 992.3674807692308, "packet_end": 1021.5554663461539, "raw_text": "其他人在浪费内存带宽和算力的环节", "selected_text": "其他人在浪费内存带宽和算力的环节", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 473 ], "packet_id": 36, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1004.3074807692309, "end": 1007.3074807692309, "text": "接下来要讲的是Deepseek在V4.1中", "raw_asr_text": "接下来要讲的是Deepseek在V4.1中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 36, "words": [ { "word": "接", "start": 1004.3074807692309, "end": 1004.6674807692308, "probability": 0.9931640625 }, { "word": "下", "start": 1004.6674807692308, "end": 1004.7874807692308, "probability": 0.99951171875 }, { "word": "来", "start": 1004.7874807692308, "end": 1004.8874807692308, "probability": 1.0 }, { "word": "要", "start": 1004.8874807692308, "end": 1005.0274807692308, "probability": 0.99365234375 }, { "word": "讲", "start": 1005.0274807692308, "end": 1005.1674807692308, "probability": 0.99853515625 }, { "word": "的是", "start": 1005.1674807692308, "end": 1005.4274807692308, "probability": 0.9873046875 }, { "word": "De", "start": 1005.4274807692308, "end": 1005.6274807692308, "probability": 0.98681640625 }, { "word": "ep", "start": 1005.6274807692308, "end": 1005.7074807692309, "probability": 0.99951171875 }, { "word": "se", "start": 1005.7074807692309, "end": 1005.8874807692308, "probability": 0.97119140625 }, { "word": "ek", "start": 1005.8874807692308, "end": 1006.0074807692308, "probability": 1.0 }, { "word": "在", "start": 1006.0074807692308, "end": 1006.3074807692309, "probability": 0.982421875 }, { "word": "V", "start": 1006.3074807692309, "end": 1006.5474807692308, "probability": 0.72119140625 }, { "word": "4.", "start": 1006.5474807692308, "end": 1006.7674807692308, "probability": 0.99267578125 }, { "word": "1", "start": 1006.8674807692308, "end": 1007.1074807692308, "probability": 0.99755859375 }, { "word": "中", "start": 1007.1074807692308, "end": 1007.3074807692309, "probability": 0.96630859375 } ], "asr_metrics": { "avg_logprob": -0.11299644470214844, "compression_ratio": 1.226775956284153, "no_speech_prob": 3.439483073353067e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "是最值得关注的到了这个阶段百万token上下文几乎已经不是这个模型最有冲击力的特性了因此Deepseek开始将注意力转向架构中其他人在浪费内存带宽和算力的环节接下来要讲的是Deepseek在V4.1中重新实现时同样做了优化的其他组件首先是MHC我之前已经做过MHC的相关讲解视频这里就长话短说V4.1针对它的专属改动叫做单路径MHC这种设计基本解决了" } ], "id": 474, "review_annotation": "", "row_id": 474, "bounded_row_id": 474, "parent_row_ids": [ 474 ], "parent_row_id": 474, "packet_id": 36, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 992.3674807692308, "packet_end": 1021.5554663461539, "raw_text": "接下来要讲的是Deepseek在V4.1中", "selected_text": "接下来要讲的是Deepseek在V4.1中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 474 ], "packet_id": 36, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1007.3074807692309, "end": 1010.1674807692308, "text": "重新实现时同样做了优化的其他组件", "raw_asr_text": "重新实现时同样做了优化的其他组件", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 36, "words": [ { "word": "重", "start": 1007.3074807692309, "end": 1007.6274807692308, "probability": 0.99169921875 }, { "word": "新", "start": 1007.6274807692308, "end": 1007.8074807692309, "probability": 0.99951171875 }, { "word": "实", "start": 1007.8074807692309, "end": 1007.9674807692309, "probability": 0.9990234375 }, { "word": "现", "start": 1007.9674807692309, "end": 1008.1474807692308, "probability": 0.9990234375 }, { "word": "时", "start": 1008.1474807692308, "end": 1008.3274807692309, "probability": 0.9970703125 }, { "word": "同", "start": 1008.3274807692309, "end": 1008.5674807692309, "probability": 0.9970703125 }, { "word": "样", "start": 1008.5674807692309, "end": 1008.7674807692308, "probability": 1.0 }, { "word": "做", "start": 1008.7674807692308, "end": 1008.9074807692308, "probability": 0.99560546875 }, { "word": "了", "start": 1008.9074807692308, "end": 1009.0474807692308, "probability": 0.99755859375 }, { "word": "优", "start": 1009.0474807692308, "end": 1009.2274807692309, "probability": 0.99951171875 }, { "word": "化", "start": 1009.2274807692309, "end": 1009.3874807692308, "probability": 1.0 }, { "word": "的", "start": 1009.3874807692308, "end": 1009.5074807692308, "probability": 0.99853515625 }, { "word": "其", "start": 1009.5074807692308, "end": 1009.7274807692309, "probability": 0.998046875 }, { "word": "他", "start": 1009.7274807692309, "end": 1009.8474807692309, "probability": 0.916015625 }, { "word": "组", "start": 1009.8474807692309, "end": 1010.0274807692308, "probability": 0.998046875 }, { "word": "件", "start": 1010.0274807692308, "end": 1010.1674807692308, "probability": 0.97998046875 } ], "asr_metrics": { "avg_logprob": -0.11299644470214844, "compression_ratio": 1.226775956284153, "no_speech_prob": 3.439483073353067e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "是最值得关注的到了这个阶段百万token上下文几乎已经不是这个模型最有冲击力的特性了因此Deepseek开始将注意力转向架构中其他人在浪费内存带宽和算力的环节接下来要讲的是Deepseek在V4.1中重新实现时同样做了优化的其他组件首先是MHC我之前已经做过MHC的相关讲解视频这里就长话短说V4.1针对它的专属改动叫做单路径MHC这种设计基本解决了" } ], "id": 475, "review_annotation": "", "row_id": 475, "bounded_row_id": 475, "parent_row_ids": [ 475 ], "parent_row_id": 475, "packet_id": 36, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 992.3674807692308, "packet_end": 1021.5554663461539, "raw_text": "重新实现时同样做了优化的其他组件", "selected_text": "重新实现时同样做了优化的其他组件", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 475 ], "packet_id": 36, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1010.5474807692308, "end": 1011.7474807692308, "text": "首先是MHC", "raw_asr_text": "首先是MHC", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 36, "words": [ { "word": "首先", "start": 1010.5474807692308, "end": 1010.9074807692308, "probability": 0.99072265625 }, { "word": "是", "start": 1010.9074807692308, "end": 1011.1074807692308, "probability": 0.9990234375 }, { "word": "M", "start": 1011.1074807692308, "end": 1011.3074807692309, "probability": 0.91943359375 }, { "word": "H", "start": 1011.3074807692309, "end": 1011.5074807692308, "probability": 0.98291015625 }, { "word": "C", "start": 1011.5074807692308, "end": 1011.7474807692308, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.11299644470214844, "compression_ratio": 1.226775956284153, "no_speech_prob": 3.439483073353067e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "是最值得关注的到了这个阶段百万token上下文几乎已经不是这个模型最有冲击力的特性了因此Deepseek开始将注意力转向架构中其他人在浪费内存带宽和算力的环节接下来要讲的是Deepseek在V4.1中重新实现时同样做了优化的其他组件首先是MHC我之前已经做过MHC的相关讲解视频这里就长话短说V4.1针对它的专属改动叫做单路径MHC这种设计基本解决了" } ], "id": 476, "review_annotation": "", "row_id": 476, "bounded_row_id": 476, "parent_row_ids": [ 476 ], "parent_row_id": 476, "packet_id": 36, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 992.3674807692308, "packet_end": 1021.5554663461539, "raw_text": "首先是MHC", "selected_text": "首先是MHC", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 476 ], "packet_id": 36, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1011.7474807692308, "end": 1014.9474807692309, "text": "我之前已经做过MHC的相关讲解视频", "raw_asr_text": "我之前已经做过MHC的相关讲解视频", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 36, "words": [ { "word": "我", "start": 1011.7474807692308, "end": 1012.0274807692308, "probability": 0.97607421875 }, { "word": "之前", "start": 1012.0274807692308, "end": 1012.3874807692308, "probability": 0.99951171875 }, { "word": "已经", "start": 1012.3874807692308, "end": 1012.6074807692308, "probability": 0.9931640625 }, { "word": "做", "start": 1012.6074807692308, "end": 1012.8474807692309, "probability": 0.998046875 }, { "word": "过", "start": 1012.8474807692309, "end": 1012.9874807692308, "probability": 0.998046875 }, { "word": "M", "start": 1012.9874807692308, "end": 1013.2674807692308, "probability": 0.99462890625 }, { "word": "H", "start": 1013.2674807692308, "end": 1013.4874807692308, "probability": 0.998046875 }, { "word": "C", "start": 1013.4874807692308, "end": 1013.6674807692308, "probability": 1.0 }, { "word": "的", "start": 1013.6674807692308, "end": 1013.8074807692309, "probability": 0.9931640625 }, { "word": "相", "start": 1013.8074807692309, "end": 1014.0074807692308, "probability": 0.99853515625 }, { "word": "关", "start": 1014.0074807692308, "end": 1014.2074807692309, "probability": 0.99951171875 }, { "word": "讲", "start": 1014.2074807692309, "end": 1014.4274807692308, "probability": 0.9970703125 }, { "word": "解", "start": 1014.4274807692308, "end": 1014.6074807692308, "probability": 0.99609375 }, { "word": "视", "start": 1014.6074807692308, "end": 1014.7674807692308, "probability": 0.998046875 }, { "word": "频", "start": 1014.7674807692308, "end": 1014.9474807692309, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11299644470214844, "compression_ratio": 1.226775956284153, "no_speech_prob": 3.439483073353067e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "是最值得关注的到了这个阶段百万token上下文几乎已经不是这个模型最有冲击力的特性了因此Deepseek开始将注意力转向架构中其他人在浪费内存带宽和算力的环节接下来要讲的是Deepseek在V4.1中重新实现时同样做了优化的其他组件首先是MHC我之前已经做过MHC的相关讲解视频这里就长话短说V4.1针对它的专属改动叫做单路径MHC这种设计基本解决了" } ], "id": 477, "review_annotation": "", "row_id": 477, "bounded_row_id": 477, "parent_row_ids": [ 477 ], "parent_row_id": 477, "packet_id": 36, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 992.3674807692308, "packet_end": 1021.5554663461539, "raw_text": "我之前已经做过MHC的相关讲解视频", "selected_text": "我之前已经做过MHC的相关讲解视频", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 477 ], "packet_id": 36, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1014.9474807692309, "end": 1016.2474807692308, "text": "这里就长话短说", "raw_asr_text": "这里就长话短说", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 36, "words": [ { "word": "这里", "start": 1014.9474807692309, "end": 1015.2074807692309, "probability": 0.9921875 }, { "word": "就", "start": 1015.2074807692309, "end": 1015.4474807692309, "probability": 0.99755859375 }, { "word": "长", "start": 1015.4474807692309, "end": 1015.6874807692309, "probability": 0.97265625 }, { "word": "话", "start": 1015.6874807692309, "end": 1015.8474807692309, "probability": 0.9990234375 }, { "word": "短", "start": 1015.8474807692309, "end": 1016.0274807692308, "probability": 0.9990234375 }, { "word": "说", "start": 1016.0274807692308, "end": 1016.2474807692308, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.11299644470214844, "compression_ratio": 1.226775956284153, "no_speech_prob": 3.439483073353067e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "是最值得关注的到了这个阶段百万token上下文几乎已经不是这个模型最有冲击力的特性了因此Deepseek开始将注意力转向架构中其他人在浪费内存带宽和算力的环节接下来要讲的是Deepseek在V4.1中重新实现时同样做了优化的其他组件首先是MHC我之前已经做过MHC的相关讲解视频这里就长话短说V4.1针对它的专属改动叫做单路径MHC这种设计基本解决了" } ], "id": 478, "review_annotation": "", "row_id": 478, "bounded_row_id": 478, "parent_row_ids": [ 478 ], "parent_row_id": 478, "packet_id": 36, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 992.3674807692308, "packet_end": 1021.5554663461539, "raw_text": "这里就长话短说", "selected_text": "这里就长话短说", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 478 ], "packet_id": 36, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1016.2474807692308, "end": 1020.1074807692308, "text": "V4.1针对它的专属改动叫做单路径MHC", "raw_asr_text": "V4.1针对它的专属改动叫做单路径MHC", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 36, "words": [ { "word": "V", "start": 1016.2474807692308, "end": 1016.7474807692308, "probability": 0.96875 }, { "word": "4.", "start": 1016.7474807692308, "end": 1016.9874807692308, "probability": 0.99755859375 }, { "word": "1", "start": 1017.1274807692308, "end": 1017.3274807692309, "probability": 0.998046875 }, { "word": "针", "start": 1017.3274807692309, "end": 1017.5674807692309, "probability": 0.998046875 }, { "word": "对", "start": 1017.5674807692309, "end": 1017.7074807692309, "probability": 0.99951171875 }, { "word": "它的", "start": 1017.7074807692309, "end": 1017.9074807692308, "probability": 0.953125 }, { "word": "专", "start": 1017.9074807692308, "end": 1018.1474807692308, "probability": 0.998046875 }, { "word": "属", "start": 1018.1474807692308, "end": 1018.3074807692309, "probability": 0.99609375 }, { "word": "改", "start": 1018.3074807692309, "end": 1018.4674807692309, "probability": 0.9970703125 }, { "word": "动", "start": 1018.4674807692309, "end": 1018.6274807692308, "probability": 0.99609375 }, { "word": "叫", "start": 1018.6274807692308, "end": 1018.8074807692309, "probability": 0.99462890625 }, { "word": "做", "start": 1018.8074807692309, "end": 1018.9674807692309, "probability": 0.98095703125 }, { "word": "单", "start": 1018.9674807692309, "end": 1019.1674807692308, "probability": 0.98779296875 }, { "word": "路", "start": 1019.1674807692308, "end": 1019.3474807692309, "probability": 0.99560546875 }, { "word": "径", "start": 1019.3474807692309, "end": 1019.4674807692309, "probability": 0.9951171875 }, { "word": "M", "start": 1019.4674807692309, "end": 1019.7074807692309, "probability": 0.974609375 }, { "word": "H", "start": 1019.7074807692309, "end": 1019.8874807692308, "probability": 0.998046875 }, { "word": "C", "start": 1019.8874807692308, "end": 1020.1074807692308, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.11299644470214844, "compression_ratio": 1.226775956284153, "no_speech_prob": 3.439483073353067e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "是最值得关注的到了这个阶段百万token上下文几乎已经不是这个模型最有冲击力的特性了因此Deepseek开始将注意力转向架构中其他人在浪费内存带宽和算力的环节接下来要讲的是Deepseek在V4.1中重新实现时同样做了优化的其他组件首先是MHC我之前已经做过MHC的相关讲解视频这里就长话短说V4.1针对它的专属改动叫做单路径MHC这种设计基本解决了" } ], "id": 479, "review_annotation": "", "row_id": 479, "bounded_row_id": 479, "parent_row_ids": [ 479 ], "parent_row_id": 479, "packet_id": 36, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 992.3674807692308, "packet_end": 1021.5554663461539, "raw_text": "V4.1针对它的专属改动叫做单路径MHC", "selected_text": "V4.1针对它的专属改动叫做单路径MHC", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 479 ], "packet_id": 36, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1020.755466346154, "end": 1022.855466346154, "text": "这种设计基本解决了原先棘手的依赖问题", "raw_asr_text": "这种设计基本解决了原先棘手的依赖问题", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 37, "words": [ { "word": "这", "start": 1019.955466346154, "end": 1020.375466346154, "probability": 0.96435546875 }, { "word": "种", "start": 1020.375466346154, "end": 1020.515466346154, "probability": 0.99755859375 }, { "word": "设", "start": 1020.515466346154, "end": 1020.695466346154, "probability": 0.998046875 }, { "word": "计", "start": 1020.695466346154, "end": 1020.7954663461541, "probability": 0.9990234375 }, { "word": "基本", "start": 1020.7954663461541, "end": 1021.095466346154, "probability": 0.998046875 }, { "word": "解", "start": 1021.095466346154, "end": 1021.355466346154, "probability": 0.99560546875 }, { "word": "决", "start": 1021.355466346154, "end": 1021.495466346154, "probability": 0.9990234375 }, { "word": "了", "start": 1021.495466346154, "end": 1021.595466346154, "probability": 0.99853515625 }, { "word": "原", "start": 1021.595466346154, "end": 1021.735466346154, "probability": 0.99609375 }, { "word": "先", "start": 1021.735466346154, "end": 1021.955466346154, "probability": 0.99755859375 }, { "word": "棘", "start": 1021.955466346154, "end": 1022.095466346154, "probability": 0.9921875 }, { "word": "手", "start": 1022.095466346154, "end": 1022.235466346154, "probability": 0.98974609375 }, { "word": "的", "start": 1022.235466346154, "end": 1022.355466346154, "probability": 0.99853515625 }, { "word": "依", "start": 1022.355466346154, "end": 1022.475466346154, "probability": 0.99951171875 }, { "word": "赖", "start": 1022.475466346154, "end": 1022.575466346154, "probability": 0.99853515625 }, { "word": "问题", "start": 1022.575466346154, "end": 1022.855466346154, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.07660435254757221, "compression_ratio": 1.256281407035176, "no_speech_prob": 3.851858762504712e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这种设计基本解决了原先棘手的依赖问题该依赖会迫使GPU多次访问残差流和原有实现相比这种新实现能将仿存流量降低约一半而且这次他们真的把Ngram落地到了模型中Deepseek此前就推出过Ngram它可以把需要记忆的信息和开销高昂的Transformer计算分离开来我也做过Ngram的相关视频但此前他们从未在自家模型中落地过这项设计这种情况直到V4.1才改变他们加入的Ngram总参数量达1960亿分布" } ], "id": 480, "review_annotation": "", "row_id": 480, "bounded_row_id": 480, "parent_row_ids": [ 480 ], "parent_row_id": 480, "packet_id": 37, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1019.955466346154, "packet_end": 1049.143451923077, "raw_text": "这种设计基本解决了原先棘手的依赖问题", "selected_text": "这种设计基本解决了原先棘手的依赖问题", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 480 ], "packet_id": 37, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1022.855466346154, "end": 1025.635466346154, "text": "该依赖会迫使GPU多次访问残差流", "raw_asr_text": "该依赖会迫使GPU多次访问残差流", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 37, "words": [ { "word": "该", "start": 1022.855466346154, "end": 1023.215466346154, "probability": 0.97265625 }, { "word": "依", "start": 1023.215466346154, "end": 1023.4154663461541, "probability": 1.0 }, { "word": "赖", "start": 1023.4154663461541, "end": 1023.5354663461541, "probability": 1.0 }, { "word": "会", "start": 1023.5354663461541, "end": 1023.735466346154, "probability": 0.99853515625 }, { "word": "迫", "start": 1023.735466346154, "end": 1023.9154663461541, "probability": 0.9990234375 }, { "word": "使", "start": 1023.9154663461541, "end": 1024.075466346154, "probability": 1.0 }, { "word": "G", "start": 1024.075466346154, "end": 1024.215466346154, "probability": 0.97900390625 }, { "word": "PU", "start": 1024.215466346154, "end": 1024.495466346154, "probability": 0.998046875 }, { "word": "多", "start": 1024.495466346154, "end": 1024.695466346154, "probability": 0.99951171875 }, { "word": "次", "start": 1024.695466346154, "end": 1024.875466346154, "probability": 1.0 }, { "word": "访", "start": 1024.875466346154, "end": 1024.995466346154, "probability": 0.998046875 }, { "word": "问", "start": 1024.995466346154, "end": 1025.175466346154, "probability": 0.99951171875 }, { "word": "残", "start": 1025.175466346154, "end": 1025.3354663461541, "probability": 0.9990234375 }, { "word": "差", "start": 1025.3354663461541, "end": 1025.475466346154, "probability": 0.7802734375 }, { "word": "流", "start": 1025.475466346154, "end": 1025.635466346154, "probability": 0.98876953125 } ], "asr_metrics": { "avg_logprob": -0.07660435254757221, "compression_ratio": 1.256281407035176, "no_speech_prob": 3.851858762504712e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这种设计基本解决了原先棘手的依赖问题该依赖会迫使GPU多次访问残差流和原有实现相比这种新实现能将仿存流量降低约一半而且这次他们真的把Ngram落地到了模型中Deepseek此前就推出过Ngram它可以把需要记忆的信息和开销高昂的Transformer计算分离开来我也做过Ngram的相关视频但此前他们从未在自家模型中落地过这项设计这种情况直到V4.1才改变他们加入的Ngram总参数量达1960亿分布" } ], "id": 481, "review_annotation": "", "row_id": 481, "bounded_row_id": 481, "parent_row_ids": [ 481 ], "parent_row_id": 481, "packet_id": 37, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1019.955466346154, "packet_end": 1049.143451923077, "raw_text": "该依赖会迫使GPU多次访问残差流", "selected_text": "该依赖会迫使GPU多次访问残差流", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 481 ], "packet_id": 37, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1025.635466346154, "end": 1026.8354663461541, "text": "和原有实现相比", "raw_asr_text": "和原有实现相比", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 37, "words": [ { "word": "和", "start": 1025.635466346154, "end": 1025.8554663461541, "probability": 0.98779296875 }, { "word": "原", "start": 1025.8554663461541, "end": 1026.015466346154, "probability": 0.99462890625 }, { "word": "有", "start": 1026.015466346154, "end": 1026.195466346154, "probability": 0.99560546875 }, { "word": "实", "start": 1026.195466346154, "end": 1026.3354663461541, "probability": 0.80029296875 }, { "word": "现", "start": 1026.3354663461541, "end": 1026.455466346154, "probability": 0.60205078125 }, { "word": "相", "start": 1026.455466346154, "end": 1026.675466346154, "probability": 0.99951171875 }, { "word": "比", "start": 1026.675466346154, "end": 1026.8354663461541, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07660435254757221, "compression_ratio": 1.256281407035176, "no_speech_prob": 3.851858762504712e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这种设计基本解决了原先棘手的依赖问题该依赖会迫使GPU多次访问残差流和原有实现相比这种新实现能将仿存流量降低约一半而且这次他们真的把Ngram落地到了模型中Deepseek此前就推出过Ngram它可以把需要记忆的信息和开销高昂的Transformer计算分离开来我也做过Ngram的相关视频但此前他们从未在自家模型中落地过这项设计这种情况直到V4.1才改变他们加入的Ngram总参数量达1960亿分布" } ], "id": 482, "review_annotation": "", "row_id": 482, "bounded_row_id": 482, "parent_row_ids": [ 482 ], "parent_row_id": 482, "packet_id": 37, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1019.955466346154, "packet_end": 1049.143451923077, "raw_text": "和原有实现相比", "selected_text": "和原有实现相比", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 482 ], "packet_id": 37, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1026.8354663461541, "end": 1029.475466346154, "text": "这种新实现能将仿存流量降低约一半", "raw_asr_text": "这种新实现能将仿存流量降低约一半", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 37, "words": [ { "word": "这", "start": 1026.8354663461541, "end": 1027.0954663461541, "probability": 0.986328125 }, { "word": "种", "start": 1027.0954663461541, "end": 1027.195466346154, "probability": 0.99951171875 }, { "word": "新", "start": 1027.195466346154, "end": 1027.395466346154, "probability": 0.99853515625 }, { "word": "实", "start": 1027.395466346154, "end": 1027.535466346154, "probability": 0.99951171875 }, { "word": "现", "start": 1027.535466346154, "end": 1027.715466346154, "probability": 0.99462890625 }, { "word": "能", "start": 1027.715466346154, "end": 1027.8554663461541, "probability": 0.9990234375 }, { "word": "将", "start": 1027.8554663461541, "end": 1028.015466346154, "probability": 0.9990234375 }, { "word": "仿", "start": 1028.015466346154, "end": 1028.215466346154, "probability": 0.9599609375 }, { "word": "存", "start": 1028.215466346154, "end": 1028.3554663461541, "probability": 0.9970703125 }, { "word": "流", "start": 1028.3554663461541, "end": 1028.515466346154, "probability": 0.99072265625 }, { "word": "量", "start": 1028.515466346154, "end": 1028.655466346154, "probability": 0.99951171875 }, { "word": "降", "start": 1028.655466346154, "end": 1028.8554663461541, "probability": 0.99853515625 }, { "word": "低", "start": 1028.8554663461541, "end": 1029.015466346154, "probability": 0.99609375 }, { "word": "约", "start": 1029.015466346154, "end": 1029.0954663461541, "probability": 0.9609375 }, { "word": "一", "start": 1029.0954663461541, "end": 1029.295466346154, "probability": 0.9970703125 }, { "word": "半", "start": 1029.295466346154, "end": 1029.475466346154, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.07660435254757221, "compression_ratio": 1.256281407035176, "no_speech_prob": 3.851858762504712e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这种设计基本解决了原先棘手的依赖问题该依赖会迫使GPU多次访问残差流和原有实现相比这种新实现能将仿存流量降低约一半而且这次他们真的把Ngram落地到了模型中Deepseek此前就推出过Ngram它可以把需要记忆的信息和开销高昂的Transformer计算分离开来我也做过Ngram的相关视频但此前他们从未在自家模型中落地过这项设计这种情况直到V4.1才改变他们加入的Ngram总参数量达1960亿分布" } ], "id": 483, "review_annotation": "", "row_id": 483, "bounded_row_id": 483, "parent_row_ids": [ 483 ], "parent_row_id": 483, "packet_id": 37, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1019.955466346154, "packet_end": 1049.143451923077, "raw_text": "这种新实现能将仿存流量降低约一半", "selected_text": "这种新实现能将仿存流量降低约一半", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 483 ], "packet_id": 37, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1029.475466346154, "end": 1032.235466346154, "text": "而且这次他们真的把Ngram落地到了模型中", "raw_asr_text": "而且这次他们真的把Ngram落地到了模型中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 37, "words": [ { "word": "而且", "start": 1029.475466346154, "end": 1029.875466346154, "probability": 0.9921875 }, { "word": "这", "start": 1029.875466346154, "end": 1030.055466346154, "probability": 0.998046875 }, { "word": "次", "start": 1030.055466346154, "end": 1030.175466346154, "probability": 0.99658203125 }, { "word": "他们", "start": 1030.175466346154, "end": 1030.375466346154, "probability": 0.9443359375 }, { "word": "真的", "start": 1030.375466346154, "end": 1030.655466346154, "probability": 0.99560546875 }, { "word": "把", "start": 1030.655466346154, "end": 1030.8554663461541, "probability": 0.9912109375 }, { "word": "N", "start": 1030.8554663461541, "end": 1030.995466346154, "probability": 0.65283203125 }, { "word": "gram", "start": 1030.995466346154, "end": 1031.215466346154, "probability": 0.892578125 }, { "word": "落", "start": 1031.215466346154, "end": 1031.395466346154, "probability": 0.99462890625 }, { "word": "地", "start": 1031.395466346154, "end": 1031.535466346154, "probability": 0.9736328125 }, { "word": "到了", "start": 1031.535466346154, "end": 1031.715466346154, "probability": 0.98046875 }, { "word": "模", "start": 1031.715466346154, "end": 1031.875466346154, "probability": 0.998046875 }, { "word": "型", "start": 1031.875466346154, "end": 1032.055466346154, "probability": 0.9990234375 }, { "word": "中", "start": 1032.055466346154, "end": 1032.235466346154, "probability": 0.9970703125 } ], "asr_metrics": { "avg_logprob": -0.07660435254757221, "compression_ratio": 1.256281407035176, "no_speech_prob": 3.851858762504712e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这种设计基本解决了原先棘手的依赖问题该依赖会迫使GPU多次访问残差流和原有实现相比这种新实现能将仿存流量降低约一半而且这次他们真的把Ngram落地到了模型中Deepseek此前就推出过Ngram它可以把需要记忆的信息和开销高昂的Transformer计算分离开来我也做过Ngram的相关视频但此前他们从未在自家模型中落地过这项设计这种情况直到V4.1才改变他们加入的Ngram总参数量达1960亿分布" } ], "id": 484, "review_annotation": "", "row_id": 484, "bounded_row_id": 484, "parent_row_ids": [ 484 ], "parent_row_id": 484, "packet_id": 37, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1019.955466346154, "packet_end": 1049.143451923077, "raw_text": "而且这次他们真的把Ngram落地到了模型中", "selected_text": "而且这次他们真的把Ngram落地到了模型中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 484 ], "packet_id": 37, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1032.235466346154, "end": 1034.115466346154, "text": "Deepseek此前就推出过Ngram", "raw_asr_text": "Deepseek此前就推出过Ngram", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 37, "words": [ { "word": "De", "start": 1032.235466346154, "end": 1032.515466346154, "probability": 0.859375 }, { "word": "ep", "start": 1032.515466346154, "end": 1032.615466346154, "probability": 0.927734375 }, { "word": "se", "start": 1032.615466346154, "end": 1032.715466346154, "probability": 0.2275390625 }, { "word": "ek", "start": 1032.715466346154, "end": 1032.815466346154, "probability": 0.779296875 }, { "word": "此", "start": 1032.815466346154, "end": 1032.955466346154, "probability": 0.98486328125 }, { "word": "前", "start": 1032.955466346154, "end": 1033.135466346154, "probability": 0.99951171875 }, { "word": "就", "start": 1033.135466346154, "end": 1033.275466346154, "probability": 0.99609375 }, { "word": "推", "start": 1033.275466346154, "end": 1033.455466346154, "probability": 0.99560546875 }, { "word": "出", "start": 1033.455466346154, "end": 1033.5954663461541, "probability": 0.99853515625 }, { "word": "过", "start": 1033.5954663461541, "end": 1033.715466346154, "probability": 0.9931640625 }, { "word": "N", "start": 1033.715466346154, "end": 1033.8554663461541, "probability": 0.98828125 }, { "word": "gram", "start": 1033.8554663461541, "end": 1034.115466346154, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.07660435254757221, "compression_ratio": 1.256281407035176, "no_speech_prob": 3.851858762504712e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这种设计基本解决了原先棘手的依赖问题该依赖会迫使GPU多次访问残差流和原有实现相比这种新实现能将仿存流量降低约一半而且这次他们真的把Ngram落地到了模型中Deepseek此前就推出过Ngram它可以把需要记忆的信息和开销高昂的Transformer计算分离开来我也做过Ngram的相关视频但此前他们从未在自家模型中落地过这项设计这种情况直到V4.1才改变他们加入的Ngram总参数量达1960亿分布" } ], "id": 485, "review_annotation": "", "row_id": 485, "bounded_row_id": 485, "parent_row_ids": [ 485 ], "parent_row_id": 485, "packet_id": 37, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1019.955466346154, "packet_end": 1049.143451923077, "raw_text": "Deepseek此前就推出过Ngram", "selected_text": "Deepseek此前就推出过Ngram", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 485 ], "packet_id": 37, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1034.115466346154, "end": 1038.415466346154, "text": "它可以把需要记忆的信息和开销高昂的Transformer计算分离开来", "raw_asr_text": "它可以把需要记忆的信息和开销高昂的Transformer计算分离开来", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 37, "words": [ { "word": "它", "start": 1034.115466346154, "end": 1034.415466346154, "probability": 0.70166015625 }, { "word": "可以", "start": 1034.415466346154, "end": 1034.655466346154, "probability": 0.99853515625 }, { "word": "把", "start": 1034.655466346154, "end": 1034.815466346154, "probability": 0.99658203125 }, { "word": "需要", "start": 1034.815466346154, "end": 1035.075466346154, "probability": 0.99755859375 }, { "word": "记", "start": 1035.075466346154, "end": 1035.3354663461541, "probability": 0.982421875 }, { "word": "忆", "start": 1035.3354663461541, "end": 1035.435466346154, "probability": 0.998046875 }, { "word": "的", "start": 1035.435466346154, "end": 1035.535466346154, "probability": 0.99853515625 }, { "word": "信", "start": 1035.535466346154, "end": 1035.695466346154, "probability": 0.99755859375 }, { "word": "息", "start": 1035.695466346154, "end": 1035.935466346154, "probability": 1.0 }, { "word": "和", "start": 1035.935466346154, "end": 1036.0954663461541, "probability": 0.99755859375 }, { "word": "开", "start": 1036.0954663461541, "end": 1036.295466346154, "probability": 0.99951171875 }, { "word": "销", "start": 1036.295466346154, "end": 1036.535466346154, "probability": 0.99853515625 }, { "word": "高", "start": 1036.535466346154, "end": 1036.715466346154, "probability": 0.9990234375 }, { "word": "昂", "start": 1036.715466346154, "end": 1036.895466346154, "probability": 0.99755859375 }, { "word": "的", "start": 1036.895466346154, "end": 1037.015466346154, "probability": 0.9990234375 }, { "word": "Trans", "start": 1037.015466346154, "end": 1037.175466346154, "probability": 0.8994140625 }, { "word": "form", "start": 1037.175466346154, "end": 1037.435466346154, "probability": 0.9990234375 }, { "word": "er", "start": 1037.435466346154, "end": 1037.535466346154, "probability": 0.9833984375 }, { "word": "计", "start": 1037.535466346154, "end": 1037.695466346154, "probability": 0.9931640625 }, { "word": "算", "start": 1037.695466346154, "end": 1037.8354663461541, "probability": 1.0 }, { "word": "分", "start": 1037.8354663461541, "end": 1037.995466346154, "probability": 0.9990234375 }, { "word": "离", "start": 1037.995466346154, "end": 1038.135466346154, "probability": 0.9970703125 }, { "word": "开", "start": 1038.135466346154, "end": 1038.255466346154, "probability": 0.9990234375 }, { "word": "来", "start": 1038.255466346154, "end": 1038.415466346154, "probability": 0.9951171875 } ], "asr_metrics": { "avg_logprob": -0.07660435254757221, "compression_ratio": 1.256281407035176, "no_speech_prob": 3.851858762504712e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这种设计基本解决了原先棘手的依赖问题该依赖会迫使GPU多次访问残差流和原有实现相比这种新实现能将仿存流量降低约一半而且这次他们真的把Ngram落地到了模型中Deepseek此前就推出过Ngram它可以把需要记忆的信息和开销高昂的Transformer计算分离开来我也做过Ngram的相关视频但此前他们从未在自家模型中落地过这项设计这种情况直到V4.1才改变他们加入的Ngram总参数量达1960亿分布" } ], "id": 486, "review_annotation": "", "row_id": 486, "bounded_row_id": 486, "parent_row_ids": [ 486 ], "parent_row_id": 486, "packet_id": 37, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1019.955466346154, "packet_end": 1049.143451923077, "raw_text": "它可以把需要记忆的信息和开销高昂的Transformer计算分离开来", "selected_text": "它可以把需要记忆的信息和开销高昂的Transformer计算分离开来", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 486 ], "packet_id": 37, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1038.415466346154, "end": 1040.255466346154, "text": "我也做过Ngram的相关视频", "raw_asr_text": "我也做过Ngram的相关视频", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 37, "words": [ { "word": "我也", "start": 1038.415466346154, "end": 1038.775466346154, "probability": 0.9892578125 }, { "word": "做", "start": 1038.775466346154, "end": 1038.955466346154, "probability": 0.9970703125 }, { "word": "过", "start": 1038.955466346154, "end": 1039.135466346154, "probability": 0.99951171875 }, { "word": "N", "start": 1039.135466346154, "end": 1039.375466346154, "probability": 0.98876953125 }, { "word": "gram", "start": 1039.375466346154, "end": 1039.575466346154, "probability": 0.99951171875 }, { "word": "的", "start": 1039.575466346154, "end": 1039.715466346154, "probability": 0.99560546875 }, { "word": "相", "start": 1039.715466346154, "end": 1039.8354663461541, "probability": 0.99853515625 }, { "word": "关", "start": 1039.8354663461541, "end": 1039.935466346154, "probability": 0.99951171875 }, { "word": "视", "start": 1039.935466346154, "end": 1040.0954663461541, "probability": 0.99951171875 }, { "word": "频", "start": 1040.0954663461541, "end": 1040.255466346154, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.07660435254757221, "compression_ratio": 1.256281407035176, "no_speech_prob": 3.851858762504712e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这种设计基本解决了原先棘手的依赖问题该依赖会迫使GPU多次访问残差流和原有实现相比这种新实现能将仿存流量降低约一半而且这次他们真的把Ngram落地到了模型中Deepseek此前就推出过Ngram它可以把需要记忆的信息和开销高昂的Transformer计算分离开来我也做过Ngram的相关视频但此前他们从未在自家模型中落地过这项设计这种情况直到V4.1才改变他们加入的Ngram总参数量达1960亿分布" } ], "id": 487, "review_annotation": "", "row_id": 487, "bounded_row_id": 487, "parent_row_ids": [ 487 ], "parent_row_id": 487, "packet_id": 37, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1019.955466346154, "packet_end": 1049.143451923077, "raw_text": "我也做过Ngram的相关视频", "selected_text": "我也做过Ngram的相关视频", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 487 ], "packet_id": 37, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1040.255466346154, "end": 1043.515466346154, "text": "但此前他们从未在自家模型中落地过这项设计", "raw_asr_text": "但此前他们从未在自家模型中落地过这项设计", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 37, "words": [ { "word": "但", "start": 1040.255466346154, "end": 1040.535466346154, "probability": 0.98876953125 }, { "word": "此", "start": 1040.535466346154, "end": 1040.755466346154, "probability": 0.99462890625 }, { "word": "前", "start": 1040.755466346154, "end": 1040.935466346154, "probability": 0.9990234375 }, { "word": "他们", "start": 1040.935466346154, "end": 1041.155466346154, "probability": 0.89404296875 }, { "word": "从", "start": 1041.155466346154, "end": 1041.3354663461541, "probability": 0.998046875 }, { "word": "未", "start": 1041.3354663461541, "end": 1041.495466346154, "probability": 0.99658203125 }, { "word": "在", "start": 1041.495466346154, "end": 1041.635466346154, "probability": 0.99755859375 }, { "word": "自", "start": 1041.635466346154, "end": 1041.795466346154, "probability": 0.9990234375 }, { "word": "家", "start": 1041.795466346154, "end": 1041.915466346154, "probability": 0.9794921875 }, { "word": "模", "start": 1041.915466346154, "end": 1042.055466346154, "probability": 0.9951171875 }, { "word": "型", "start": 1042.055466346154, "end": 1042.195466346154, "probability": 0.99951171875 }, { "word": "中", "start": 1042.195466346154, "end": 1042.3354663461541, "probability": 0.998046875 }, { "word": "落", "start": 1042.3354663461541, "end": 1042.575466346154, "probability": 0.99853515625 }, { "word": "地", "start": 1042.575466346154, "end": 1042.755466346154, "probability": 0.99365234375 }, { "word": "过", "start": 1042.755466346154, "end": 1042.915466346154, "probability": 0.99853515625 }, { "word": "这", "start": 1042.915466346154, "end": 1043.075466346154, "probability": 0.998046875 }, { "word": "项", "start": 1043.075466346154, "end": 1043.175466346154, "probability": 0.998046875 }, { "word": "设", "start": 1043.175466346154, "end": 1043.3554663461541, "probability": 1.0 }, { "word": "计", "start": 1043.3554663461541, "end": 1043.515466346154, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07660435254757221, "compression_ratio": 1.256281407035176, "no_speech_prob": 3.851858762504712e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这种设计基本解决了原先棘手的依赖问题该依赖会迫使GPU多次访问残差流和原有实现相比这种新实现能将仿存流量降低约一半而且这次他们真的把Ngram落地到了模型中Deepseek此前就推出过Ngram它可以把需要记忆的信息和开销高昂的Transformer计算分离开来我也做过Ngram的相关视频但此前他们从未在自家模型中落地过这项设计这种情况直到V4.1才改变他们加入的Ngram总参数量达1960亿分布" } ], "id": 488, "review_annotation": "", "row_id": 488, "bounded_row_id": 488, "parent_row_ids": [ 488 ], "parent_row_id": 488, "packet_id": 37, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1019.955466346154, "packet_end": 1049.143451923077, "raw_text": "但此前他们从未在自家模型中落地过这项设计", "selected_text": "但此前他们从未在自家模型中落地过这项设计", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 488 ], "packet_id": 37, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1043.515466346154, "end": 1045.815466346154, "text": "这种情况直到V4.1才改变", "raw_asr_text": "这种情况直到V4.1才改变", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 37, "words": [ { "word": "这", "start": 1043.515466346154, "end": 1043.8354663461541, "probability": 0.99755859375 }, { "word": "种", "start": 1043.8354663461541, "end": 1043.955466346154, "probability": 0.99951171875 }, { "word": "情况", "start": 1043.955466346154, "end": 1044.235466346154, "probability": 0.9990234375 }, { "word": "直", "start": 1044.235466346154, "end": 1044.475466346154, "probability": 0.99853515625 }, { "word": "到", "start": 1044.475466346154, "end": 1044.695466346154, "probability": 0.99951171875 }, { "word": "V", "start": 1044.695466346154, "end": 1044.8354663461541, "probability": 0.88818359375 }, { "word": "4.", "start": 1044.8354663461541, "end": 1045.035466346154, "probability": 0.990234375 }, { "word": "1", "start": 1045.155466346154, "end": 1045.315466346154, "probability": 0.99951171875 }, { "word": "才", "start": 1045.315466346154, "end": 1045.455466346154, "probability": 0.98876953125 }, { "word": "改", "start": 1045.455466346154, "end": 1045.635466346154, "probability": 0.99951171875 }, { "word": "变", "start": 1045.635466346154, "end": 1045.815466346154, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07660435254757221, "compression_ratio": 1.256281407035176, "no_speech_prob": 3.851858762504712e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这种设计基本解决了原先棘手的依赖问题该依赖会迫使GPU多次访问残差流和原有实现相比这种新实现能将仿存流量降低约一半而且这次他们真的把Ngram落地到了模型中Deepseek此前就推出过Ngram它可以把需要记忆的信息和开销高昂的Transformer计算分离开来我也做过Ngram的相关视频但此前他们从未在自家模型中落地过这项设计这种情况直到V4.1才改变他们加入的Ngram总参数量达1960亿分布" } ], "id": 489, "review_annotation": "", "row_id": 489, "bounded_row_id": 489, "parent_row_ids": [ 489 ], "parent_row_id": 489, "packet_id": 37, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1019.955466346154, "packet_end": 1049.143451923077, "raw_text": "这种情况直到V4.1才改变", "selected_text": "这种情况直到V4.1才改变", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 489 ], "packet_id": 37, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1045.815466346154, "end": 1048.343451923077, "text": "他们加入的Ngram总参数量达1960亿分布", "raw_asr_text": "他们加入的Ngram总参数量达1960亿分布", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 37, "words": [ { "word": "他们", "start": 1045.815466346154, "end": 1046.135466346154, "probability": 0.89599609375 }, { "word": "加入", "start": 1046.135466346154, "end": 1046.375466346154, "probability": 0.99951171875 }, { "word": "的", "start": 1046.375466346154, "end": 1046.515466346154, "probability": 0.9951171875 }, { "word": "N", "start": 1046.515466346154, "end": 1046.695466346154, "probability": 0.99609375 }, { "word": "gram", "start": 1046.695466346154, "end": 1046.915466346154, "probability": 0.99951171875 }, { "word": "总", "start": 1046.915466346154, "end": 1047.0954663461541, "probability": 0.99658203125 }, { "word": "参", "start": 1047.0954663461541, "end": 1047.275466346154, "probability": 0.9931640625 }, { "word": "数", "start": 1047.275466346154, "end": 1047.415466346154, "probability": 0.99951171875 }, { "word": "量", "start": 1047.415466346154, "end": 1047.535466346154, "probability": 0.998046875 }, { "word": "达", "start": 1047.535466346154, "end": 1047.755466346154, "probability": 0.9990234375 }, { "word": "19", "start": 1047.755466346154, "end": 1048.015466346154, "probability": 0.990234375 }, { "word": "60", "start": 1048.015466346154, "end": 1048.635466346154, "probability": 0.998046875 }, { "word": "亿", "start": 1048.635466346154, "end": 1048.8354663461541, "probability": 0.9970703125 }, { "word": "分", "start": 1048.8354663461541, "end": 1049.035466346154, "probability": 0.98876953125 }, { "word": "布", "start": 1049.035466346154, "end": 1049.115466346154, "probability": 0.685546875 } ], "asr_metrics": { "avg_logprob": -0.07660435254757221, "compression_ratio": 1.256281407035176, "no_speech_prob": 3.851858762504712e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这种设计基本解决了原先棘手的依赖问题该依赖会迫使GPU多次访问残差流和原有实现相比这种新实现能将仿存流量降低约一半而且这次他们真的把Ngram落地到了模型中Deepseek此前就推出过Ngram它可以把需要记忆的信息和开销高昂的Transformer计算分离开来我也做过Ngram的相关视频但此前他们从未在自家模型中落地过这项设计这种情况直到V4.1才改变他们加入的Ngram总参数量达1960亿分布" } ], "id": 490, "review_annotation": "", "row_id": 490, "bounded_row_id": 490, "parent_row_ids": [ 490 ], "parent_row_id": 490, "packet_id": 37, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1019.955466346154, "packet_end": 1049.143451923077, "raw_text": "他们加入的Ngram总参数量达1960亿分布", "selected_text": "他们加入的Ngram总参数量达1960亿分布", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 490 ], "packet_id": 37, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1048.343451923077, "end": 1050.103451923077, "text": "达1960亿分布在两个模块中", "raw_asr_text": "达1960亿分布在两个模块中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 38, "words": [ { "word": "达", "start": 1047.543451923077, "end": 1047.7434519230771, "probability": 0.666015625 }, { "word": "19", "start": 1047.7434519230771, "end": 1048.0234519230771, "probability": 0.95068359375 }, { "word": "60", "start": 1048.0234519230771, "end": 1048.643451923077, "probability": 0.998046875 }, { "word": "亿", "start": 1048.643451923077, "end": 1048.823451923077, "probability": 0.998046875 }, { "word": "分", "start": 1048.823451923077, "end": 1049.043451923077, "probability": 0.89501953125 }, { "word": "布", "start": 1049.043451923077, "end": 1049.143451923077, "probability": 0.974609375 }, { "word": "在", "start": 1049.143451923077, "end": 1049.2434519230771, "probability": 0.9951171875 }, { "word": "两", "start": 1049.2434519230771, "end": 1049.4434519230772, "probability": 0.96484375 }, { "word": "个", "start": 1049.4434519230772, "end": 1049.623451923077, "probability": 0.998046875 }, { "word": "模", "start": 1049.623451923077, "end": 1049.7434519230771, "probability": 0.99462890625 }, { "word": "块", "start": 1049.7434519230771, "end": 1049.9234519230772, "probability": 0.99951171875 }, { "word": "中", "start": 1049.9234519230772, "end": 1050.103451923077, "probability": 0.984375 } ], "asr_metrics": { "avg_logprob": -0.09243850901647267, "compression_ratio": 1.2789968652037618, "no_speech_prob": 3.988839120117049e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "达1960亿分布在两个模块中所以从技术层面来说该模型总共存储的参数量约为5520亿加1960亿合计7480亿但这多出来的1960亿参数是通过哈希Ngram查照表进行悉数访问的条件记忆而非每个Token都会激活的另外1960亿个稠密参数也就是说Ngram大幅提升了模型的记忆容量但它单Token所需的算力仍和5520亿参数规模的MOE模型相当最后一会儿" } ], "id": 491, "review_annotation": "", "row_id": 491, "bounded_row_id": 491, "parent_row_ids": [ 491 ], "parent_row_id": 491, "packet_id": 38, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1047.543451923077, "packet_end": 1076.7314374999999, "raw_text": "达1960亿分布在两个模块中", "selected_text": "达1960亿分布在两个模块中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 491 ], "packet_id": 38, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1050.103451923077, "end": 1051.383451923077, "text": "所以从技术层面来说", "raw_asr_text": "所以从技术层面来说", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 38, "words": [ { "word": "所以", "start": 1050.103451923077, "end": 1050.4234519230772, "probability": 0.96923828125 }, { "word": "从", "start": 1050.4234519230772, "end": 1050.583451923077, "probability": 0.99365234375 }, { "word": "技", "start": 1050.583451923077, "end": 1050.783451923077, "probability": 0.9990234375 }, { "word": "术", "start": 1050.783451923077, "end": 1050.883451923077, "probability": 1.0 }, { "word": "层", "start": 1050.883451923077, "end": 1050.9634519230772, "probability": 0.9970703125 }, { "word": "面", "start": 1050.9634519230772, "end": 1051.103451923077, "probability": 0.9892578125 }, { "word": "来", "start": 1051.103451923077, "end": 1051.2034519230772, "probability": 0.99755859375 }, { "word": "说", "start": 1051.2034519230772, "end": 1051.383451923077, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.09243850901647267, "compression_ratio": 1.2789968652037618, "no_speech_prob": 3.988839120117049e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "达1960亿分布在两个模块中所以从技术层面来说该模型总共存储的参数量约为5520亿加1960亿合计7480亿但这多出来的1960亿参数是通过哈希Ngram查照表进行悉数访问的条件记忆而非每个Token都会激活的另外1960亿个稠密参数也就是说Ngram大幅提升了模型的记忆容量但它单Token所需的算力仍和5520亿参数规模的MOE模型相当最后一会儿" } ], "id": 492, "review_annotation": "", "row_id": 492, "bounded_row_id": 492, "parent_row_ids": [ 492 ], "parent_row_id": 492, "packet_id": 38, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1047.543451923077, "packet_end": 1076.7314374999999, "raw_text": "所以从技术层面来说", "selected_text": "所以从技术层面来说", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 492 ], "packet_id": 38, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1051.383451923077, "end": 1053.343451923077, "text": "该模型总共存储的参数量", "raw_asr_text": "该模型总共存储的参数量", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 38, "words": [ { "word": "该", "start": 1051.383451923077, "end": 1051.623451923077, "probability": 0.97265625 }, { "word": "模", "start": 1051.623451923077, "end": 1051.7634519230771, "probability": 0.99951171875 }, { "word": "型", "start": 1051.7634519230771, "end": 1051.9434519230772, "probability": 0.99951171875 }, { "word": "总", "start": 1051.9434519230772, "end": 1052.163451923077, "probability": 0.99365234375 }, { "word": "共", "start": 1052.163451923077, "end": 1052.323451923077, "probability": 0.9970703125 }, { "word": "存", "start": 1052.323451923077, "end": 1052.563451923077, "probability": 0.99951171875 }, { "word": "储", "start": 1052.563451923077, "end": 1052.7234519230772, "probability": 0.9990234375 }, { "word": "的", "start": 1052.7234519230772, "end": 1052.843451923077, "probability": 0.9990234375 }, { "word": "参", "start": 1052.843451923077, "end": 1053.043451923077, "probability": 0.9970703125 }, { "word": "数", "start": 1053.043451923077, "end": 1053.1834519230772, "probability": 0.99951171875 }, { "word": "量", "start": 1053.1834519230772, "end": 1053.343451923077, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.09243850901647267, "compression_ratio": 1.2789968652037618, "no_speech_prob": 3.988839120117049e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "达1960亿分布在两个模块中所以从技术层面来说该模型总共存储的参数量约为5520亿加1960亿合计7480亿但这多出来的1960亿参数是通过哈希Ngram查照表进行悉数访问的条件记忆而非每个Token都会激活的另外1960亿个稠密参数也就是说Ngram大幅提升了模型的记忆容量但它单Token所需的算力仍和5520亿参数规模的MOE模型相当最后一会儿" } ], "id": 493, "review_annotation": "", "row_id": 493, "bounded_row_id": 493, "parent_row_ids": [ 493 ], "parent_row_id": 493, "packet_id": 38, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1047.543451923077, "packet_end": 1076.7314374999999, "raw_text": "该模型总共存储的参数量", "selected_text": "该模型总共存储的参数量", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 493 ], "packet_id": 38, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1053.343451923077, "end": 1055.883451923077, "text": "约为5520亿加1960亿", "raw_asr_text": "约为5520亿加1960亿", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 38, "words": [ { "word": "约", "start": 1053.343451923077, "end": 1053.543451923077, "probability": 0.9970703125 }, { "word": "为", "start": 1053.543451923077, "end": 1053.7634519230771, "probability": 0.99658203125 }, { "word": "55", "start": 1053.7634519230771, "end": 1054.123451923077, "probability": 0.95947265625 }, { "word": "20", "start": 1054.123451923077, "end": 1054.7234519230772, "probability": 0.9990234375 }, { "word": "亿", "start": 1054.7234519230772, "end": 1054.883451923077, "probability": 0.99609375 }, { "word": "加", "start": 1054.883451923077, "end": 1055.043451923077, "probability": 0.65966796875 }, { "word": "19", "start": 1055.043451923077, "end": 1055.303451923077, "probability": 0.99658203125 }, { "word": "60", "start": 1055.303451923077, "end": 1055.7034519230772, "probability": 0.99951171875 }, { "word": "亿", "start": 1055.7034519230772, "end": 1055.883451923077, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09243850901647267, "compression_ratio": 1.2789968652037618, "no_speech_prob": 3.988839120117049e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "达1960亿分布在两个模块中所以从技术层面来说该模型总共存储的参数量约为5520亿加1960亿合计7480亿但这多出来的1960亿参数是通过哈希Ngram查照表进行悉数访问的条件记忆而非每个Token都会激活的另外1960亿个稠密参数也就是说Ngram大幅提升了模型的记忆容量但它单Token所需的算力仍和5520亿参数规模的MOE模型相当最后一会儿" } ], "id": 494, "review_annotation": "", "row_id": 494, "bounded_row_id": 494, "parent_row_ids": [ 494 ], "parent_row_id": 494, "packet_id": 38, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1047.543451923077, "packet_end": 1076.7314374999999, "raw_text": "约为5520亿加1960亿", "selected_text": "约为5520亿加1960亿", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 494 ], "packet_id": 38, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1055.883451923077, "end": 1057.4434519230772, "text": "合计7480亿", "raw_asr_text": "合计7480亿", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 38, "words": [ { "word": "合", "start": 1055.883451923077, "end": 1056.2434519230771, "probability": 0.97802734375 }, { "word": "计", "start": 1056.2434519230771, "end": 1056.383451923077, "probability": 0.99072265625 }, { "word": "74", "start": 1056.383451923077, "end": 1056.783451923077, "probability": 0.9228515625 }, { "word": "80", "start": 1056.783451923077, "end": 1057.2234519230772, "probability": 0.99951171875 }, { "word": "亿", "start": 1057.2234519230772, "end": 1057.4434519230772, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09243850901647267, "compression_ratio": 1.2789968652037618, "no_speech_prob": 3.988839120117049e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "达1960亿分布在两个模块中所以从技术层面来说该模型总共存储的参数量约为5520亿加1960亿合计7480亿但这多出来的1960亿参数是通过哈希Ngram查照表进行悉数访问的条件记忆而非每个Token都会激活的另外1960亿个稠密参数也就是说Ngram大幅提升了模型的记忆容量但它单Token所需的算力仍和5520亿参数规模的MOE模型相当最后一会儿" } ], "id": 495, "review_annotation": "", "row_id": 495, "bounded_row_id": 495, "parent_row_ids": [ 495 ], "parent_row_id": 495, "packet_id": 38, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1047.543451923077, "packet_end": 1076.7314374999999, "raw_text": "合计7480亿", "selected_text": "合计7480亿", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 495 ], "packet_id": 38, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1057.4434519230772, "end": 1059.9434519230772, "text": "但这多出来的1960亿参数", "raw_asr_text": "但这多出来的1960亿参数", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 38, "words": [ { "word": "但", "start": 1057.4434519230772, "end": 1057.623451923077, "probability": 0.96630859375 }, { "word": "这", "start": 1057.623451923077, "end": 1057.863451923077, "probability": 0.98876953125 }, { "word": "多", "start": 1057.863451923077, "end": 1058.163451923077, "probability": 0.9951171875 }, { "word": "出来", "start": 1058.163451923077, "end": 1058.4834519230772, "probability": 0.9970703125 }, { "word": "的", "start": 1058.4834519230772, "end": 1058.6834519230772, "probability": 0.9990234375 }, { "word": "19", "start": 1058.6834519230772, "end": 1058.903451923077, "probability": 0.98974609375 }, { "word": "60", "start": 1058.903451923077, "end": 1059.4634519230772, "probability": 0.9990234375 }, { "word": "亿", "start": 1059.4634519230772, "end": 1059.623451923077, "probability": 1.0 }, { "word": "参", "start": 1059.623451923077, "end": 1059.783451923077, "probability": 0.9951171875 }, { "word": "数", "start": 1059.783451923077, "end": 1059.9434519230772, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09243850901647267, "compression_ratio": 1.2789968652037618, "no_speech_prob": 3.988839120117049e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "达1960亿分布在两个模块中所以从技术层面来说该模型总共存储的参数量约为5520亿加1960亿合计7480亿但这多出来的1960亿参数是通过哈希Ngram查照表进行悉数访问的条件记忆而非每个Token都会激活的另外1960亿个稠密参数也就是说Ngram大幅提升了模型的记忆容量但它单Token所需的算力仍和5520亿参数规模的MOE模型相当最后一会儿" } ], "id": 496, "review_annotation": "", "row_id": 496, "bounded_row_id": 496, "parent_row_ids": [ 496 ], "parent_row_id": 496, "packet_id": 38, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1047.543451923077, "packet_end": 1076.7314374999999, "raw_text": "但这多出来的1960亿参数", "selected_text": "但这多出来的1960亿参数", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 496 ], "packet_id": 38, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1059.9434519230772, "end": 1061.9634519230772, "text": "是通过哈希Ngram查照表", "raw_asr_text": "是通过哈希Ngram查照表", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 38, "words": [ { "word": "是", "start": 1059.9434519230772, "end": 1060.083451923077, "probability": 0.99609375 }, { "word": "通", "start": 1060.083451923077, "end": 1060.2634519230771, "probability": 0.99365234375 }, { "word": "过", "start": 1060.2634519230771, "end": 1060.4234519230772, "probability": 0.99951171875 }, { "word": "哈", "start": 1060.4234519230772, "end": 1060.7034519230772, "probability": 0.7333984375 }, { "word": "希", "start": 1060.7034519230772, "end": 1060.9434519230772, "probability": 0.86669921875 }, { "word": "N", "start": 1060.9434519230772, "end": 1061.2034519230772, "probability": 0.6015625 }, { "word": "gram", "start": 1061.2034519230772, "end": 1061.4234519230772, "probability": 0.9326171875 }, { "word": "查", "start": 1061.4234519230772, "end": 1061.663451923077, "probability": 0.92431640625 }, { "word": "照", "start": 1061.663451923077, "end": 1061.823451923077, "probability": 0.85791015625 }, { "word": "表", "start": 1061.823451923077, "end": 1061.9634519230772, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.09243850901647267, "compression_ratio": 1.2789968652037618, "no_speech_prob": 3.988839120117049e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "达1960亿分布在两个模块中所以从技术层面来说该模型总共存储的参数量约为5520亿加1960亿合计7480亿但这多出来的1960亿参数是通过哈希Ngram查照表进行悉数访问的条件记忆而非每个Token都会激活的另外1960亿个稠密参数也就是说Ngram大幅提升了模型的记忆容量但它单Token所需的算力仍和5520亿参数规模的MOE模型相当最后一会儿" } ], "id": 497, "review_annotation": "", "row_id": 497, "bounded_row_id": 497, "parent_row_ids": [ 497 ], "parent_row_id": 497, "packet_id": 38, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1047.543451923077, "packet_end": 1076.7314374999999, "raw_text": "是通过哈希Ngram查照表", "selected_text": "是通过哈希Ngram查照表", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 497 ], "packet_id": 38, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1061.9634519230772, "end": 1063.7634519230771, "text": "进行悉数访问的条件记忆", "raw_asr_text": "进行悉数访问的条件记忆", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 38, "words": [ { "word": "进", "start": 1061.9634519230772, "end": 1062.2034519230772, "probability": 0.99609375 }, { "word": "行", "start": 1062.2034519230772, "end": 1062.403451923077, "probability": 0.99951171875 }, { "word": "悉", "start": 1062.403451923077, "end": 1062.563451923077, "probability": 0.6796875 }, { "word": "数", "start": 1062.563451923077, "end": 1062.7634519230771, "probability": 0.56689453125 }, { "word": "访", "start": 1062.7634519230771, "end": 1062.9234519230772, "probability": 0.9990234375 }, { "word": "问", "start": 1062.9234519230772, "end": 1063.043451923077, "probability": 0.99951171875 }, { "word": "的", "start": 1063.043451923077, "end": 1063.143451923077, "probability": 0.9921875 }, { "word": "条", "start": 1063.143451923077, "end": 1063.283451923077, "probability": 0.99853515625 }, { "word": "件", "start": 1063.283451923077, "end": 1063.4434519230772, "probability": 1.0 }, { "word": "记", "start": 1063.4434519230772, "end": 1063.603451923077, "probability": 0.98583984375 }, { "word": "忆", "start": 1063.603451923077, "end": 1063.7634519230771, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.09243850901647267, "compression_ratio": 1.2789968652037618, "no_speech_prob": 3.988839120117049e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "达1960亿分布在两个模块中所以从技术层面来说该模型总共存储的参数量约为5520亿加1960亿合计7480亿但这多出来的1960亿参数是通过哈希Ngram查照表进行悉数访问的条件记忆而非每个Token都会激活的另外1960亿个稠密参数也就是说Ngram大幅提升了模型的记忆容量但它单Token所需的算力仍和5520亿参数规模的MOE模型相当最后一会儿" } ], "id": 498, "review_annotation": "", "row_id": 498, "bounded_row_id": 498, "parent_row_ids": [ 498 ], "parent_row_id": 498, "packet_id": 38, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1047.543451923077, "packet_end": 1076.7314374999999, "raw_text": "进行悉数访问的条件记忆", "selected_text": "进行悉数访问的条件记忆", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 498 ], "packet_id": 38, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1063.7634519230771, "end": 1066.2234519230772, "text": "而非每个Token都会激活的", "raw_asr_text": "而非每个Token都会激活的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 38, "words": [ { "word": "而", "start": 1063.7634519230771, "end": 1064.363451923077, "probability": 0.9970703125 }, { "word": "非", "start": 1064.363451923077, "end": 1064.543451923077, "probability": 0.99951171875 }, { "word": "每", "start": 1064.543451923077, "end": 1064.7034519230772, "probability": 0.99951171875 }, { "word": "个", "start": 1064.7034519230772, "end": 1064.843451923077, "probability": 0.9970703125 }, { "word": "T", "start": 1064.843451923077, "end": 1064.9634519230772, "probability": 0.6474609375 }, { "word": "oken", "start": 1064.9634519230772, "end": 1065.163451923077, "probability": 0.98193359375 }, { "word": "都", "start": 1065.163451923077, "end": 1065.383451923077, "probability": 0.9853515625 }, { "word": "会", "start": 1065.383451923077, "end": 1065.603451923077, "probability": 0.9990234375 }, { "word": "激", "start": 1065.603451923077, "end": 1065.803451923077, "probability": 0.998046875 }, { "word": "活", "start": 1065.803451923077, "end": 1066.043451923077, "probability": 0.99951171875 }, { "word": "的", "start": 1066.043451923077, "end": 1066.2234519230772, "probability": 0.994140625 } ], "asr_metrics": { "avg_logprob": -0.09243850901647267, "compression_ratio": 1.2789968652037618, "no_speech_prob": 3.988839120117049e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "达1960亿分布在两个模块中所以从技术层面来说该模型总共存储的参数量约为5520亿加1960亿合计7480亿但这多出来的1960亿参数是通过哈希Ngram查照表进行悉数访问的条件记忆而非每个Token都会激活的另外1960亿个稠密参数也就是说Ngram大幅提升了模型的记忆容量但它单Token所需的算力仍和5520亿参数规模的MOE模型相当最后一会儿" } ], "id": 499, "review_annotation": "", "row_id": 499, "bounded_row_id": 499, "parent_row_ids": [ 499 ], "parent_row_id": 499, "packet_id": 38, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1047.543451923077, "packet_end": 1076.7314374999999, "raw_text": "而非每个Token都会激活的", "selected_text": "而非每个Token都会激活的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 499 ], "packet_id": 38, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1066.2234519230772, "end": 1068.403451923077, "text": "另外1960亿个稠密参数", "raw_asr_text": "另外1960亿个稠密参数", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 38, "words": [ { "word": "另外", "start": 1066.2234519230772, "end": 1066.543451923077, "probability": 0.99267578125 }, { "word": "19", "start": 1066.543451923077, "end": 1066.803451923077, "probability": 0.9951171875 }, { "word": "60", "start": 1066.803451923077, "end": 1067.4634519230772, "probability": 0.927734375 }, { "word": "亿", "start": 1067.4634519230772, "end": 1067.663451923077, "probability": 0.962890625 }, { "word": "个", "start": 1067.663451923077, "end": 1067.783451923077, "probability": 0.9921875 }, { "word": "稠", "start": 1067.783451923077, "end": 1067.9634519230772, "probability": 0.94140625 }, { "word": "密", "start": 1067.9634519230772, "end": 1068.083451923077, "probability": 0.599609375 }, { "word": "参", "start": 1068.083451923077, "end": 1068.2434519230771, "probability": 0.99560546875 }, { "word": "数", "start": 1068.2434519230771, "end": 1068.403451923077, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09243850901647267, "compression_ratio": 1.2789968652037618, "no_speech_prob": 3.988839120117049e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "达1960亿分布在两个模块中所以从技术层面来说该模型总共存储的参数量约为5520亿加1960亿合计7480亿但这多出来的1960亿参数是通过哈希Ngram查照表进行悉数访问的条件记忆而非每个Token都会激活的另外1960亿个稠密参数也就是说Ngram大幅提升了模型的记忆容量但它单Token所需的算力仍和5520亿参数规模的MOE模型相当最后一会儿" } ], "id": 500, "review_annotation": "", "row_id": 500, "bounded_row_id": 500, "parent_row_ids": [ 500 ], "parent_row_id": 500, "packet_id": 38, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1047.543451923077, "packet_end": 1076.7314374999999, "raw_text": "另外1960亿个稠密参数", "selected_text": "另外1960亿个稠密参数", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 500 ], "packet_id": 38, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1068.403451923077, "end": 1070.303451923077, "text": "也就是说Ngram大幅提升了", "raw_asr_text": "也就是说Ngram大幅提升了", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 38, "words": [ { "word": "也", "start": 1068.403451923077, "end": 1068.7034519230772, "probability": 0.9970703125 }, { "word": "就是", "start": 1068.7034519230772, "end": 1068.843451923077, "probability": 0.8642578125 }, { "word": "说", "start": 1068.843451923077, "end": 1068.9834519230772, "probability": 0.94775390625 }, { "word": "N", "start": 1068.9834519230772, "end": 1069.123451923077, "probability": 0.95068359375 }, { "word": "gram", "start": 1069.123451923077, "end": 1069.363451923077, "probability": 0.99951171875 }, { "word": "大", "start": 1069.363451923077, "end": 1069.663451923077, "probability": 0.9931640625 }, { "word": "幅", "start": 1069.663451923077, "end": 1069.863451923077, "probability": 0.9990234375 }, { "word": "提", "start": 1069.863451923077, "end": 1070.0034519230771, "probability": 0.9970703125 }, { "word": "升", "start": 1070.0034519230771, "end": 1070.1834519230772, "probability": 0.9970703125 }, { "word": "了", "start": 1070.1834519230772, "end": 1070.303451923077, "probability": 0.994140625 } ], "asr_metrics": { "avg_logprob": -0.09243850901647267, "compression_ratio": 1.2789968652037618, "no_speech_prob": 3.988839120117049e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "达1960亿分布在两个模块中所以从技术层面来说该模型总共存储的参数量约为5520亿加1960亿合计7480亿但这多出来的1960亿参数是通过哈希Ngram查照表进行悉数访问的条件记忆而非每个Token都会激活的另外1960亿个稠密参数也就是说Ngram大幅提升了模型的记忆容量但它单Token所需的算力仍和5520亿参数规模的MOE模型相当最后一会儿" } ], "id": 501, "review_annotation": "", "row_id": 501, "bounded_row_id": 501, "parent_row_ids": [ 501 ], "parent_row_id": 501, "packet_id": 38, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1047.543451923077, "packet_end": 1076.7314374999999, "raw_text": "也就是说Ngram大幅提升了", "selected_text": "也就是说Ngram大幅提升了", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 501 ], "packet_id": 38, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1070.303451923077, "end": 1071.403451923077, "text": "模型的记忆容量", "raw_asr_text": "模型的记忆容量", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 38, "words": [ { "word": "模", "start": 1070.303451923077, "end": 1070.4634519230772, "probability": 0.99609375 }, { "word": "型", "start": 1070.4634519230772, "end": 1070.623451923077, "probability": 0.99951171875 }, { "word": "的", "start": 1070.623451923077, "end": 1070.783451923077, "probability": 0.998046875 }, { "word": "记", "start": 1070.783451923077, "end": 1070.9434519230772, "probability": 0.9951171875 }, { "word": "忆", "start": 1070.9434519230772, "end": 1071.063451923077, "probability": 1.0 }, { "word": "容", "start": 1071.063451923077, "end": 1071.1834519230772, "probability": 0.99755859375 }, { "word": "量", "start": 1071.1834519230772, "end": 1071.403451923077, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09243850901647267, "compression_ratio": 1.2789968652037618, "no_speech_prob": 3.988839120117049e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "达1960亿分布在两个模块中所以从技术层面来说该模型总共存储的参数量约为5520亿加1960亿合计7480亿但这多出来的1960亿参数是通过哈希Ngram查照表进行悉数访问的条件记忆而非每个Token都会激活的另外1960亿个稠密参数也就是说Ngram大幅提升了模型的记忆容量但它单Token所需的算力仍和5520亿参数规模的MOE模型相当最后一会儿" } ], "id": 502, "review_annotation": "", "row_id": 502, "bounded_row_id": 502, "parent_row_ids": [ 502 ], "parent_row_id": 502, "packet_id": 38, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1047.543451923077, "packet_end": 1076.7314374999999, "raw_text": "模型的记忆容量", "selected_text": "模型的记忆容量", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 502 ], "packet_id": 38, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1071.403451923077, "end": 1073.163451923077, "text": "但它单Token所需的算力", "raw_asr_text": "但它单Token所需的算力", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 38, "words": [ { "word": "但", "start": 1071.403451923077, "end": 1071.783451923077, "probability": 0.99462890625 }, { "word": "它", "start": 1071.783451923077, "end": 1071.9434519230772, "probability": 0.80517578125 }, { "word": "单", "start": 1071.9434519230772, "end": 1072.103451923077, "probability": 0.998046875 }, { "word": "T", "start": 1072.103451923077, "end": 1072.2634519230771, "probability": 0.93603515625 }, { "word": "oken", "start": 1072.2634519230771, "end": 1072.4234519230772, "probability": 0.99853515625 }, { "word": "所", "start": 1072.4234519230772, "end": 1072.603451923077, "probability": 0.9697265625 }, { "word": "需", "start": 1072.603451923077, "end": 1072.7234519230772, "probability": 0.99755859375 }, { "word": "的", "start": 1072.7234519230772, "end": 1072.823451923077, "probability": 0.998046875 }, { "word": "算", "start": 1072.823451923077, "end": 1072.9834519230772, "probability": 0.9990234375 }, { "word": "力", "start": 1072.9834519230772, "end": 1073.163451923077, "probability": 0.990234375 } ], "asr_metrics": { "avg_logprob": -0.09243850901647267, "compression_ratio": 1.2789968652037618, "no_speech_prob": 3.988839120117049e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "达1960亿分布在两个模块中所以从技术层面来说该模型总共存储的参数量约为5520亿加1960亿合计7480亿但这多出来的1960亿参数是通过哈希Ngram查照表进行悉数访问的条件记忆而非每个Token都会激活的另外1960亿个稠密参数也就是说Ngram大幅提升了模型的记忆容量但它单Token所需的算力仍和5520亿参数规模的MOE模型相当最后一会儿" } ], "id": 503, "review_annotation": "", "row_id": 503, "bounded_row_id": 503, "parent_row_ids": [ 503 ], "parent_row_id": 503, "packet_id": 38, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1047.543451923077, "packet_end": 1076.7314374999999, "raw_text": "但它单Token所需的算力", "selected_text": "但它单Token所需的算力", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 503 ], "packet_id": 38, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1073.163451923077, "end": 1075.4234519230772, "text": "仍和5520亿参数规模的", "raw_asr_text": "仍和5520亿参数规模的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 38, "words": [ { "word": "仍", "start": 1073.163451923077, "end": 1073.603451923077, "probability": 0.9970703125 }, { "word": "和", "start": 1073.603451923077, "end": 1073.803451923077, "probability": 0.994140625 }, { "word": "55", "start": 1073.803451923077, "end": 1074.0234519230771, "probability": 0.9375 }, { "word": "20", "start": 1074.0234519230771, "end": 1074.643451923077, "probability": 0.9990234375 }, { "word": "亿", "start": 1074.643451923077, "end": 1074.783451923077, "probability": 1.0 }, { "word": "参", "start": 1074.783451923077, "end": 1074.9434519230772, "probability": 0.9951171875 }, { "word": "数", "start": 1074.9434519230772, "end": 1075.063451923077, "probability": 1.0 }, { "word": "规", "start": 1075.063451923077, "end": 1075.2034519230772, "probability": 0.998046875 }, { "word": "模", "start": 1075.2034519230772, "end": 1075.323451923077, "probability": 0.9970703125 }, { "word": "的", "start": 1075.323451923077, "end": 1075.4234519230772, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.09243850901647267, "compression_ratio": 1.2789968652037618, "no_speech_prob": 3.988839120117049e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "达1960亿分布在两个模块中所以从技术层面来说该模型总共存储的参数量约为5520亿加1960亿合计7480亿但这多出来的1960亿参数是通过哈希Ngram查照表进行悉数访问的条件记忆而非每个Token都会激活的另外1960亿个稠密参数也就是说Ngram大幅提升了模型的记忆容量但它单Token所需的算力仍和5520亿参数规模的MOE模型相当最后一会儿" } ], "id": 504, "review_annotation": "", "row_id": 504, "bounded_row_id": 504, "parent_row_ids": [ 504 ], "parent_row_id": 504, "packet_id": 38, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1047.543451923077, "packet_end": 1076.7314374999999, "raw_text": "仍和5520亿参数规模的", "selected_text": "仍和5520亿参数规模的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 504 ], "packet_id": 38, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1075.4234519230772, "end": 1075.9314375, "text": "MOE模型相当", "raw_asr_text": "MOE模型相当", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 38, "words": [ { "word": "MO", "start": 1075.4234519230772, "end": 1075.7034519230772, "probability": 0.80126953125 }, { "word": "E", "start": 1075.7034519230772, "end": 1075.803451923077, "probability": 0.7646484375 }, { "word": "模", "start": 1075.803451923077, "end": 1075.903451923077, "probability": 0.99658203125 }, { "word": "型", "start": 1075.903451923077, "end": 1076.043451923077, "probability": 0.99951171875 }, { "word": "相", "start": 1076.043451923077, "end": 1076.2234519230772, "probability": 0.99560546875 }, { "word": "当", "start": 1076.2234519230772, "end": 1076.383451923077, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.09243850901647267, "compression_ratio": 1.2789968652037618, "no_speech_prob": 3.988839120117049e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "达1960亿分布在两个模块中所以从技术层面来说该模型总共存储的参数量约为5520亿加1960亿合计7480亿但这多出来的1960亿参数是通过哈希Ngram查照表进行悉数访问的条件记忆而非每个Token都会激活的另外1960亿个稠密参数也就是说Ngram大幅提升了模型的记忆容量但它单Token所需的算力仍和5520亿参数规模的MOE模型相当最后一会儿" } ], "id": 505, "review_annotation": "", "row_id": 505, "bounded_row_id": 505, "parent_row_ids": [ 505 ], "parent_row_id": 505, "packet_id": 38, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1047.543451923077, "packet_end": 1076.7314374999999, "raw_text": "MOE模型相当", "selected_text": "MOE模型相当", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 505 ], "packet_id": 38, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1076.3314375, "end": 1078.1514375, "text": "最后要介绍的是Spark", "raw_asr_text": "最后要介绍的是Spark", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 39, "words": [ { "word": "最", "start": 1076.3314375, "end": 1076.6914375, "probability": 0.75439453125 }, { "word": "后", "start": 1076.6914375, "end": 1076.9314375, "probability": 0.98974609375 }, { "word": "要", "start": 1076.9314375, "end": 1077.1114375, "probability": 0.9677734375 }, { "word": "介", "start": 1077.1114375, "end": 1077.2714375, "probability": 0.99755859375 }, { "word": "绍", "start": 1077.2714375, "end": 1077.4114375, "probability": 0.99951171875 }, { "word": "的是", "start": 1077.4114375, "end": 1077.6714375, "probability": 0.98046875 }, { "word": "Sp", "start": 1077.6714375, "end": 1077.9514374999999, "probability": 0.5673828125 }, { "word": "ark", "start": 1077.9514374999999, "end": 1078.1514375, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.11193140426484666, "compression_ratio": 1.245014245014245, "no_speech_prob": 6.271724806161671e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最后要介绍的是Spark我也做过关于Spark的视频这里简单概括他们用Spark实现了更快的解码速度且无需将草稿模型整合到主干网络的预训练流程中但这次的架构升级并不只局限于文本领域V4.1的另一项重大升级是Deep Seek终于让模型实现了原生多模态他们从零训练了自研视觉编码器Deep Seek VIT随后采用3x3像素合并加混洗的方式压缩视觉特别" } ], "id": 506, "review_annotation": "", "row_id": 506, "bounded_row_id": 506, "parent_row_ids": [ 506 ], "parent_row_id": 506, "packet_id": 39, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1075.1314375, "packet_end": 1104.3194230769232, "raw_text": "最后要介绍的是Spark", "selected_text": "最后要介绍的是Spark", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 506 ], "packet_id": 39, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1078.1514375, "end": 1080.1714375, "text": "我也做过关于Spark的视频", "raw_asr_text": "我也做过关于Spark的视频", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 39, "words": [ { "word": "我也", "start": 1078.1514375, "end": 1078.6714375, "probability": 0.939453125 }, { "word": "做", "start": 1078.6714375, "end": 1078.8714375, "probability": 0.99365234375 }, { "word": "过", "start": 1078.8714375, "end": 1079.0314375, "probability": 0.99755859375 }, { "word": "关", "start": 1079.0314375, "end": 1079.1714375, "probability": 0.99169921875 }, { "word": "于", "start": 1079.1714375, "end": 1079.2914375, "probability": 0.982421875 }, { "word": "Sp", "start": 1079.2914375, "end": 1079.4714374999999, "probability": 0.92578125 }, { "word": "ark", "start": 1079.4714374999999, "end": 1079.6514375, "probability": 0.998046875 }, { "word": "的", "start": 1079.6514375, "end": 1079.8314375, "probability": 0.97021484375 }, { "word": "视", "start": 1079.8314375, "end": 1079.9914374999998, "probability": 0.99951171875 }, { "word": "频", "start": 1079.9914374999998, "end": 1080.1714375, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.11193140426484666, "compression_ratio": 1.245014245014245, "no_speech_prob": 6.271724806161671e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最后要介绍的是Spark我也做过关于Spark的视频这里简单概括他们用Spark实现了更快的解码速度且无需将草稿模型整合到主干网络的预训练流程中但这次的架构升级并不只局限于文本领域V4.1的另一项重大升级是Deep Seek终于让模型实现了原生多模态他们从零训练了自研视觉编码器Deep Seek VIT随后采用3x3像素合并加混洗的方式压缩视觉特别" } ], "id": 507, "review_annotation": "", "row_id": 507, "bounded_row_id": 507, "parent_row_ids": [ 507 ], "parent_row_id": 507, "packet_id": 39, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1075.1314375, "packet_end": 1104.3194230769232, "raw_text": "我也做过关于Spark的视频", "selected_text": "我也做过关于Spark的视频", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 507 ], "packet_id": 39, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1080.1714375, "end": 1081.3914375, "text": "这里简单概括", "raw_asr_text": "这里简单概括", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 39, "words": [ { "word": "这里", "start": 1080.1714375, "end": 1080.4314375, "probability": 0.9677734375 }, { "word": "简", "start": 1080.4314375, "end": 1080.8314375, "probability": 0.9951171875 }, { "word": "单", "start": 1080.8314375, "end": 1081.0114375, "probability": 1.0 }, { "word": "概", "start": 1081.0114375, "end": 1081.1514375, "probability": 0.99951171875 }, { "word": "括", "start": 1081.1514375, "end": 1081.3914375, "probability": 0.9970703125 } ], "asr_metrics": { "avg_logprob": -0.11193140426484666, "compression_ratio": 1.245014245014245, "no_speech_prob": 6.271724806161671e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最后要介绍的是Spark我也做过关于Spark的视频这里简单概括他们用Spark实现了更快的解码速度且无需将草稿模型整合到主干网络的预训练流程中但这次的架构升级并不只局限于文本领域V4.1的另一项重大升级是Deep Seek终于让模型实现了原生多模态他们从零训练了自研视觉编码器Deep Seek VIT随后采用3x3像素合并加混洗的方式压缩视觉特别" } ], "id": 508, "review_annotation": "", "row_id": 508, "bounded_row_id": 508, "parent_row_ids": [ 508 ], "parent_row_id": 508, "packet_id": 39, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1075.1314375, "packet_end": 1104.3194230769232, "raw_text": "这里简单概括", "selected_text": "这里简单概括", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 508 ], "packet_id": 39, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1081.3914375, "end": 1083.7914375, "text": "他们用Spark实现了更快的解码速度", "raw_asr_text": "他们用Spark实现了更快的解码速度", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 39, "words": [ { "word": "他们", "start": 1081.3914375, "end": 1081.6914375, "probability": 0.8486328125 }, { "word": "用", "start": 1081.6914375, "end": 1081.8314375, "probability": 0.9951171875 }, { "word": "Sp", "start": 1081.8314375, "end": 1082.0114375, "probability": 0.96533203125 }, { "word": "ark", "start": 1082.0114375, "end": 1082.2114374999999, "probability": 0.9990234375 }, { "word": "实", "start": 1082.2114374999999, "end": 1082.4314375, "probability": 0.97900390625 }, { "word": "现", "start": 1082.4314375, "end": 1082.5914375, "probability": 0.9990234375 }, { "word": "了", "start": 1082.5914375, "end": 1082.6914375, "probability": 0.99072265625 }, { "word": "更", "start": 1082.6914375, "end": 1082.8514375, "probability": 0.9970703125 }, { "word": "快", "start": 1082.8514375, "end": 1083.0314375, "probability": 0.99853515625 }, { "word": "的", "start": 1083.0314375, "end": 1083.1514375, "probability": 0.99267578125 }, { "word": "解", "start": 1083.1514375, "end": 1083.2714375, "probability": 0.990234375 }, { "word": "码", "start": 1083.2714375, "end": 1083.4514374999999, "probability": 0.998046875 }, { "word": "速", "start": 1083.4514374999999, "end": 1083.6114375, "probability": 0.99853515625 }, { "word": "度", "start": 1083.6114375, "end": 1083.7914375, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11193140426484666, "compression_ratio": 1.245014245014245, "no_speech_prob": 6.271724806161671e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最后要介绍的是Spark我也做过关于Spark的视频这里简单概括他们用Spark实现了更快的解码速度且无需将草稿模型整合到主干网络的预训练流程中但这次的架构升级并不只局限于文本领域V4.1的另一项重大升级是Deep Seek终于让模型实现了原生多模态他们从零训练了自研视觉编码器Deep Seek VIT随后采用3x3像素合并加混洗的方式压缩视觉特别" } ], "id": 509, "review_annotation": "", "row_id": 509, "bounded_row_id": 509, "parent_row_ids": [ 509 ], "parent_row_id": 509, "packet_id": 39, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1075.1314375, "packet_end": 1104.3194230769232, "raw_text": "他们用Spark实现了更快的解码速度", "selected_text": "他们用Spark实现了更快的解码速度", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 509 ], "packet_id": 39, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1083.7914375, "end": 1087.5914375, "text": "且无需将草稿模型整合到主干网络的预训练流程中", "raw_asr_text": "且无需将草稿模型整合到主干网络的预训练流程中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 39, "words": [ { "word": "且", "start": 1083.7914375, "end": 1084.1114375, "probability": 0.96044921875 }, { "word": "无", "start": 1084.1114375, "end": 1084.3114375, "probability": 0.9970703125 }, { "word": "需", "start": 1084.3114375, "end": 1084.4714374999999, "probability": 0.9755859375 }, { "word": "将", "start": 1084.4714374999999, "end": 1084.6514375, "probability": 0.99658203125 }, { "word": "草", "start": 1084.6514375, "end": 1084.8514375, "probability": 0.98779296875 }, { "word": "稿", "start": 1084.8514375, "end": 1084.9714374999999, "probability": 0.9970703125 }, { "word": "模", "start": 1084.9714374999999, "end": 1085.1114375, "probability": 0.9990234375 }, { "word": "型", "start": 1085.1114375, "end": 1085.3114375, "probability": 0.99853515625 }, { "word": "整", "start": 1085.3114375, "end": 1085.4914374999998, "probability": 0.99658203125 }, { "word": "合", "start": 1085.4914374999998, "end": 1085.6914375, "probability": 0.9990234375 }, { "word": "到", "start": 1085.6914375, "end": 1085.8514375, "probability": 0.9931640625 }, { "word": "主", "start": 1085.8514375, "end": 1086.0514375, "probability": 0.998046875 }, { "word": "干", "start": 1086.0514375, "end": 1086.1714375, "probability": 0.98193359375 }, { "word": "网", "start": 1086.1714375, "end": 1086.3714375, "probability": 0.998046875 }, { "word": "络", "start": 1086.3714375, "end": 1086.5114375, "probability": 0.99853515625 }, { "word": "的", "start": 1086.5114375, "end": 1086.6514375, "probability": 0.9951171875 }, { "word": "预", "start": 1086.6514375, "end": 1086.8114375, "probability": 0.99267578125 }, { "word": "训", "start": 1086.8114375, "end": 1086.9914374999998, "probability": 0.998046875 }, { "word": "练", "start": 1086.9914374999998, "end": 1087.1114375, "probability": 0.99609375 }, { "word": "流", "start": 1087.1114375, "end": 1087.2514374999998, "probability": 0.96533203125 }, { "word": "程", "start": 1087.2514374999998, "end": 1087.4314375, "probability": 0.99951171875 }, { "word": "中", "start": 1087.4314375, "end": 1087.5914375, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.11193140426484666, "compression_ratio": 1.245014245014245, "no_speech_prob": 6.271724806161671e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最后要介绍的是Spark我也做过关于Spark的视频这里简单概括他们用Spark实现了更快的解码速度且无需将草稿模型整合到主干网络的预训练流程中但这次的架构升级并不只局限于文本领域V4.1的另一项重大升级是Deep Seek终于让模型实现了原生多模态他们从零训练了自研视觉编码器Deep Seek VIT随后采用3x3像素合并加混洗的方式压缩视觉特别" } ], "id": 510, "review_annotation": "", "row_id": 510, "bounded_row_id": 510, "parent_row_ids": [ 510 ], "parent_row_id": 510, "packet_id": 39, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1075.1314375, "packet_end": 1104.3194230769232, "raw_text": "且无需将草稿模型整合到主干网络的预训练流程中", "selected_text": "且无需将草稿模型整合到主干网络的预训练流程中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 510 ], "packet_id": 39, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1087.5914375, "end": 1090.8714375, "text": "但这次的架构升级并不只局限于文本领域", "raw_asr_text": "但这次的架构升级并不只局限于文本领域", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 39, "words": [ { "word": "但", "start": 1087.5914375, "end": 1087.8514375, "probability": 0.98583984375 }, { "word": "这", "start": 1087.8514375, "end": 1088.1114375, "probability": 0.99169921875 }, { "word": "次", "start": 1088.1114375, "end": 1088.2514374999998, "probability": 0.99755859375 }, { "word": "的", "start": 1088.2514374999998, "end": 1088.3514375, "probability": 0.89306640625 }, { "word": "架", "start": 1088.3514375, "end": 1088.5114375, "probability": 0.9990234375 }, { "word": "构", "start": 1088.5114375, "end": 1088.6714375, "probability": 0.9990234375 }, { "word": "升", "start": 1088.6714375, "end": 1088.7914375, "probability": 0.99560546875 }, { "word": "级", "start": 1088.7914375, "end": 1089.0714375, "probability": 0.9990234375 }, { "word": "并", "start": 1089.0714375, "end": 1089.3514375, "probability": 0.9912109375 }, { "word": "不", "start": 1089.3514375, "end": 1089.4714374999999, "probability": 0.99609375 }, { "word": "只", "start": 1089.4714374999999, "end": 1089.6714375, "probability": 0.86865234375 }, { "word": "局", "start": 1089.6714375, "end": 1089.7914375, "probability": 0.91796875 }, { "word": "限", "start": 1089.7914375, "end": 1089.9514374999999, "probability": 0.99853515625 }, { "word": "于", "start": 1089.9514374999999, "end": 1090.1314375, "probability": 0.99365234375 }, { "word": "文", "start": 1090.1314375, "end": 1090.3714375, "probability": 0.99072265625 }, { "word": "本", "start": 1090.3714375, "end": 1090.5514375, "probability": 0.994140625 }, { "word": "领", "start": 1090.5514375, "end": 1090.6914375, "probability": 1.0 }, { "word": "域", "start": 1090.6914375, "end": 1090.8714375, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.11193140426484666, "compression_ratio": 1.245014245014245, "no_speech_prob": 6.271724806161671e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最后要介绍的是Spark我也做过关于Spark的视频这里简单概括他们用Spark实现了更快的解码速度且无需将草稿模型整合到主干网络的预训练流程中但这次的架构升级并不只局限于文本领域V4.1的另一项重大升级是Deep Seek终于让模型实现了原生多模态他们从零训练了自研视觉编码器Deep Seek VIT随后采用3x3像素合并加混洗的方式压缩视觉特别" } ], "id": 511, "review_annotation": "", "row_id": 511, "bounded_row_id": 511, "parent_row_ids": [ 511 ], "parent_row_id": 511, "packet_id": 39, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1075.1314375, "packet_end": 1104.3194230769232, "raw_text": "但这次的架构升级并不只局限于文本领域", "selected_text": "但这次的架构升级并不只局限于文本领域", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 511 ], "packet_id": 39, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1090.8714375, "end": 1093.6914375, "text": "V4.1的另一项重大升级是Deep Seek", "raw_asr_text": "V4.1的另一项重大升级是Deep Seek", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 39, "words": [ { "word": "V", "start": 1090.8714375, "end": 1091.1914375, "probability": 0.79052734375 }, { "word": "4.", "start": 1091.1914375, "end": 1091.3514375, "probability": 0.99072265625 }, { "word": "1", "start": 1091.4714374999999, "end": 1091.6314375, "probability": 0.9970703125 }, { "word": "的", "start": 1091.6314375, "end": 1091.7714375, "probability": 0.9638671875 }, { "word": "另", "start": 1091.7714375, "end": 1091.9314375, "probability": 0.99462890625 }, { "word": "一", "start": 1091.9314375, "end": 1092.0514375, "probability": 0.994140625 }, { "word": "项", "start": 1092.0514375, "end": 1092.1514375, "probability": 0.9990234375 }, { "word": "重", "start": 1092.1514375, "end": 1092.2914375, "probability": 0.9970703125 }, { "word": "大", "start": 1092.2914375, "end": 1092.4714374999999, "probability": 0.998046875 }, { "word": "升", "start": 1092.4714374999999, "end": 1092.6314375, "probability": 0.9990234375 }, { "word": "级", "start": 1092.6314375, "end": 1092.8314375, "probability": 1.0 }, { "word": "是", "start": 1092.8314375, "end": 1093.0514375, "probability": 0.98388671875 }, { "word": "De", "start": 1093.0514375, "end": 1093.3114375, "probability": 0.93505859375 }, { "word": "ep", "start": 1093.3114375, "end": 1093.4114375, "probability": 0.97216796875 }, { "word": " Se", "start": 1093.4114375, "end": 1093.5514375, "probability": 0.387939453125 }, { "word": "ek", "start": 1093.5514375, "end": 1093.6914375, "probability": 0.97119140625 } ], "asr_metrics": { "avg_logprob": -0.11193140426484666, "compression_ratio": 1.245014245014245, "no_speech_prob": 6.271724806161671e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最后要介绍的是Spark我也做过关于Spark的视频这里简单概括他们用Spark实现了更快的解码速度且无需将草稿模型整合到主干网络的预训练流程中但这次的架构升级并不只局限于文本领域V4.1的另一项重大升级是Deep Seek终于让模型实现了原生多模态他们从零训练了自研视觉编码器Deep Seek VIT随后采用3x3像素合并加混洗的方式压缩视觉特别" } ], "id": 512, "review_annotation": "", "row_id": 512, "bounded_row_id": 512, "parent_row_ids": [ 512 ], "parent_row_id": 512, "packet_id": 39, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1075.1314375, "packet_end": 1104.3194230769232, "raw_text": "V4.1的另一项重大升级是Deep Seek", "selected_text": "V4.1的另一项重大升级是Deep Seek", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 512 ], "packet_id": 39, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1093.6914375, "end": 1096.1914375, "text": "终于让模型实现了原生多模态", "raw_asr_text": "终于让模型实现了原生多模态", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 39, "words": [ { "word": "终", "start": 1093.6914375, "end": 1094.0314375, "probability": 0.9873046875 }, { "word": "于", "start": 1094.0314375, "end": 1094.1514375, "probability": 0.99853515625 }, { "word": "让", "start": 1094.1514375, "end": 1094.2914375, "probability": 0.9990234375 }, { "word": "模", "start": 1094.2914375, "end": 1094.4714374999999, "probability": 0.99853515625 }, { "word": "型", "start": 1094.4714374999999, "end": 1094.6714375, "probability": 0.99951171875 }, { "word": "实", "start": 1094.6714375, "end": 1094.8514375, "probability": 0.99658203125 }, { "word": "现", "start": 1094.8514375, "end": 1095.0314375, "probability": 0.99951171875 }, { "word": "了", "start": 1095.0314375, "end": 1095.1514375, "probability": 0.9931640625 }, { "word": "原", "start": 1095.1514375, "end": 1095.3514375, "probability": 0.6630859375 }, { "word": "生", "start": 1095.3514375, "end": 1095.6114375, "probability": 0.9169921875 }, { "word": "多", "start": 1095.6114375, "end": 1095.7914375, "probability": 0.99560546875 }, { "word": "模", "start": 1095.7914375, "end": 1095.9314375, "probability": 0.343505859375 }, { "word": "态", "start": 1095.9314375, "end": 1096.1914375, "probability": 0.9892578125 } ], "asr_metrics": { "avg_logprob": -0.11193140426484666, "compression_ratio": 1.245014245014245, "no_speech_prob": 6.271724806161671e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最后要介绍的是Spark我也做过关于Spark的视频这里简单概括他们用Spark实现了更快的解码速度且无需将草稿模型整合到主干网络的预训练流程中但这次的架构升级并不只局限于文本领域V4.1的另一项重大升级是Deep Seek终于让模型实现了原生多模态他们从零训练了自研视觉编码器Deep Seek VIT随后采用3x3像素合并加混洗的方式压缩视觉特别" } ], "id": 513, "review_annotation": "", "row_id": 513, "bounded_row_id": 513, "parent_row_ids": [ 513 ], "parent_row_id": 513, "packet_id": 39, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1075.1314375, "packet_end": 1104.3194230769232, "raw_text": "终于让模型实现了原生多模态", "selected_text": "终于让模型实现了原生多模态", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 513 ], "packet_id": 39, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1096.1914375, "end": 1100.0514375, "text": "他们从零训练了自研视觉编码器Deep Seek VIT", "raw_asr_text": "他们从零训练了自研视觉编码器Deep Seek VIT", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 39, "words": [ { "word": "他们", "start": 1096.1914375, "end": 1096.6714375, "probability": 0.97509765625 }, { "word": "从", "start": 1096.6714375, "end": 1096.9714374999999, "probability": 0.6748046875 }, { "word": "零", "start": 1096.9714374999999, "end": 1097.1914375, "probability": 0.85009765625 }, { "word": "训", "start": 1097.1914375, "end": 1097.4114375, "probability": 0.998046875 }, { "word": "练", "start": 1097.4114375, "end": 1097.5714375, "probability": 0.99951171875 }, { "word": "了", "start": 1097.5714375, "end": 1097.7114374999999, "probability": 0.99658203125 }, { "word": "自", "start": 1097.7114374999999, "end": 1097.9314375, "probability": 0.9033203125 }, { "word": "研", "start": 1097.9314375, "end": 1098.1314375, "probability": 0.97216796875 }, { "word": "视", "start": 1098.1314375, "end": 1098.3114375, "probability": 0.99755859375 }, { "word": "觉", "start": 1098.3114375, "end": 1098.4514374999999, "probability": 0.99658203125 }, { "word": "编", "start": 1098.4514374999999, "end": 1098.5514375, "probability": 0.9716796875 }, { "word": "码", "start": 1098.5514375, "end": 1098.7514374999998, "probability": 0.9990234375 }, { "word": "器", "start": 1098.7514374999998, "end": 1098.9514374999999, "probability": 0.99951171875 }, { "word": "De", "start": 1098.9514374999999, "end": 1099.1914375, "probability": 0.99072265625 }, { "word": "ep", "start": 1099.1914375, "end": 1099.2914375, "probability": 0.99853515625 }, { "word": " Se", "start": 1099.2914375, "end": 1099.4314375, "probability": 0.9814453125 }, { "word": "ek", "start": 1099.4314375, "end": 1099.5514375, "probability": 1.0 }, { "word": " V", "start": 1099.5514375, "end": 1099.7714375, "probability": 0.609375 }, { "word": "IT", "start": 1099.7714375, "end": 1100.0514375, "probability": 0.94921875 } ], "asr_metrics": { "avg_logprob": -0.11193140426484666, "compression_ratio": 1.245014245014245, "no_speech_prob": 6.271724806161671e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最后要介绍的是Spark我也做过关于Spark的视频这里简单概括他们用Spark实现了更快的解码速度且无需将草稿模型整合到主干网络的预训练流程中但这次的架构升级并不只局限于文本领域V4.1的另一项重大升级是Deep Seek终于让模型实现了原生多模态他们从零训练了自研视觉编码器Deep Seek VIT随后采用3x3像素合并加混洗的方式压缩视觉特别" } ], "id": 514, "review_annotation": "", "row_id": 514, "bounded_row_id": 514, "parent_row_ids": [ 514 ], "parent_row_id": 514, "packet_id": 39, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1075.1314375, "packet_end": 1104.3194230769232, "raw_text": "他们从零训练了自研视觉编码器Deep Seek VIT", "selected_text": "他们从零训练了自研视觉编码器Deep Seek VIT", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 514 ], "packet_id": 39, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1100.0514375, "end": 1103.5194230769232, "text": "随后采用3x3像素合并加混洗的方式压缩视觉特别", "raw_asr_text": "随后采用3x3像素合并加混洗的方式压缩视觉特别", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 39, "words": [ { "word": "随", "start": 1100.0514375, "end": 1100.5314375, "probability": 0.978515625 }, { "word": "后", "start": 1100.5314375, "end": 1100.6514375, "probability": 0.99951171875 }, { "word": "采", "start": 1100.6514375, "end": 1100.8714375, "probability": 0.998046875 }, { "word": "用", "start": 1100.8714375, "end": 1101.0314375, "probability": 0.99951171875 }, { "word": "3", "start": 1101.0314375, "end": 1101.2714375, "probability": 0.56103515625 }, { "word": "x", "start": 1101.2714375, "end": 1101.3914375, "probability": 0.41845703125 }, { "word": "3", "start": 1101.3914375, "end": 1101.6114375, "probability": 0.9775390625 }, { "word": "像", "start": 1101.6114375, "end": 1101.7314374999999, "probability": 0.93798828125 }, { "word": "素", "start": 1101.7314374999999, "end": 1101.9914374999998, "probability": 1.0 }, { "word": "合", "start": 1101.9914374999998, "end": 1102.2514374999998, "probability": 0.98828125 }, { "word": "并", "start": 1102.2514374999998, "end": 1102.4314375, "probability": 0.9873046875 }, { "word": "加", "start": 1102.4314375, "end": 1102.6314375, "probability": 0.9755859375 }, { "word": "混", "start": 1102.6314375, "end": 1102.8514375, "probability": 0.9921875 }, { "word": "洗", "start": 1102.8514375, "end": 1103.0514375, "probability": 0.68310546875 }, { "word": "的", "start": 1103.0514375, "end": 1103.1514375, "probability": 0.982421875 }, { "word": "方", "start": 1103.1514375, "end": 1103.3114375, "probability": 0.99853515625 }, { "word": "式", "start": 1103.3114375, "end": 1103.4914374999998, "probability": 0.99951171875 }, { "word": "压", "start": 1103.4914374999998, "end": 1103.7314374999999, "probability": 0.9970703125 }, { "word": "缩", "start": 1103.7314374999999, "end": 1103.9114375, "probability": 0.9990234375 }, { "word": "视", "start": 1103.9114375, "end": 1104.1114375, "probability": 0.99609375 }, { "word": "觉", "start": 1104.1114375, "end": 1104.2314374999999, "probability": 0.99853515625 }, { "word": "特", "start": 1104.2314374999999, "end": 1104.2914375, "probability": 0.330322265625 }, { "word": "别", "start": 1104.2914375, "end": 1104.2914375, "probability": 0.301025390625 } ], "asr_metrics": { "avg_logprob": -0.11193140426484666, "compression_ratio": 1.245014245014245, "no_speech_prob": 6.271724806161671e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最后要介绍的是Spark我也做过关于Spark的视频这里简单概括他们用Spark实现了更快的解码速度且无需将草稿模型整合到主干网络的预训练流程中但这次的架构升级并不只局限于文本领域V4.1的另一项重大升级是Deep Seek终于让模型实现了原生多模态他们从零训练了自研视觉编码器Deep Seek VIT随后采用3x3像素合并加混洗的方式压缩视觉特别" } ], "id": 515, "review_annotation": "", "row_id": 515, "bounded_row_id": 515, "parent_row_ids": [ 515 ], "parent_row_id": 515, "packet_id": 39, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1075.1314375, "packet_end": 1104.3194230769232, "raw_text": "随后采用3x3像素合并加混洗的方式压缩视觉特别", "selected_text": "随后采用3x3像素合并加混洗的方式压缩视觉特别", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 515 ], "packet_id": 39, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1103.5194230769232, "end": 1104.5394230769232, "text": "在混洗的方式压缩视觉特征", "raw_asr_text": "在混洗的方式压缩视觉特征", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "在", "start": 1102.7194230769232, "end": 1102.7194230769232, "probability": 0.0256500244140625 }, { "word": "混", "start": 1102.7194230769232, "end": 1102.8194230769232, "probability": 0.4140625 }, { "word": "洗", "start": 1102.8194230769232, "end": 1103.0594230769232, "probability": 0.572265625 }, { "word": "的", "start": 1103.0594230769232, "end": 1103.1594230769233, "probability": 0.9853515625 }, { "word": "方", "start": 1103.1594230769233, "end": 1103.2994230769232, "probability": 0.99560546875 }, { "word": "式", "start": 1103.2994230769232, "end": 1103.4794230769232, "probability": 0.99951171875 }, { "word": "压", "start": 1103.4794230769232, "end": 1103.7194230769232, "probability": 0.98046875 }, { "word": "缩", "start": 1103.7194230769232, "end": 1103.8994230769233, "probability": 0.9970703125 }, { "word": "视", "start": 1103.8994230769233, "end": 1104.1194230769233, "probability": 0.990234375 }, { "word": "觉", "start": 1104.1194230769233, "end": 1104.2194230769232, "probability": 0.99853515625 }, { "word": "特", "start": 1104.2194230769232, "end": 1104.3394230769231, "probability": 0.998046875 }, { "word": "征", "start": 1104.3394230769231, "end": 1104.5394230769232, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.1282342020670573, "compression_ratio": 1.3015873015873016, "no_speech_prob": 3.882096033525073e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 516, "review_annotation": "", "row_id": 516, "bounded_row_id": 516, "parent_row_ids": [ 516 ], "parent_row_id": 516, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "在混洗的方式压缩视觉特征", "selected_text": "在混洗的方式压缩视觉特征", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 516 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1104.5394230769232, "end": 1106.3194230769232, "text": "再将其输入语言模型", "raw_asr_text": "再将其输入语言模型", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "再", "start": 1104.5394230769232, "end": 1104.7994230769232, "probability": 0.9248046875 }, { "word": "将", "start": 1104.7994230769232, "end": 1105.0194230769232, "probability": 0.9931640625 }, { "word": "其", "start": 1105.0194230769232, "end": 1105.1994230769233, "probability": 0.99853515625 }, { "word": "输", "start": 1105.1994230769233, "end": 1105.4594230769233, "probability": 0.9990234375 }, { "word": "入", "start": 1105.4594230769233, "end": 1105.5994230769234, "probability": 1.0 }, { "word": "语", "start": 1105.5994230769234, "end": 1105.7994230769232, "probability": 0.9658203125 }, { "word": "言", "start": 1105.7994230769232, "end": 1105.9594230769233, "probability": 0.9990234375 }, { "word": "模", "start": 1105.9594230769233, "end": 1106.1194230769233, "probability": 0.99853515625 }, { "word": "型", "start": 1106.1194230769233, "end": 1106.3194230769232, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.1282342020670573, "compression_ratio": 1.3015873015873016, "no_speech_prob": 3.882096033525073e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 517, "review_annotation": "", "row_id": 517, "bounded_row_id": 517, "parent_row_ids": [ 517 ], "parent_row_id": 517, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "再将其输入语言模型", "selected_text": "再将其输入语言模型", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 517 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1106.3194230769232, "end": 1108.5594230769232, "text": "这一操作将视觉tokens的数量", "raw_asr_text": "这一操作将视觉tokens的数量", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "这", "start": 1106.3194230769232, "end": 1106.5394230769232, "probability": 0.97021484375 }, { "word": "一", "start": 1106.5394230769232, "end": 1106.6794230769233, "probability": 0.99462890625 }, { "word": "操", "start": 1106.6794230769233, "end": 1106.8794230769233, "probability": 0.99951171875 }, { "word": "作", "start": 1106.8794230769233, "end": 1107.0194230769232, "probability": 0.99853515625 }, { "word": "将", "start": 1107.0194230769232, "end": 1107.2194230769232, "probability": 0.99609375 }, { "word": "视", "start": 1107.2194230769232, "end": 1107.4594230769233, "probability": 0.9970703125 }, { "word": "觉", "start": 1107.4594230769233, "end": 1107.6194230769233, "probability": 0.99853515625 }, { "word": "t", "start": 1107.6194230769233, "end": 1107.7794230769232, "probability": 0.55029296875 }, { "word": "ok", "start": 1107.7794230769232, "end": 1107.8994230769233, "probability": 0.71923828125 }, { "word": "ens", "start": 1107.8994230769233, "end": 1108.0594230769232, "probability": 0.9755859375 }, { "word": "的", "start": 1108.0594230769232, "end": 1108.1994230769233, "probability": 0.9365234375 }, { "word": "数", "start": 1108.1994230769233, "end": 1108.3994230769233, "probability": 0.99755859375 }, { "word": "量", "start": 1108.3994230769233, "end": 1108.5594230769232, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1282342020670573, "compression_ratio": 1.3015873015873016, "no_speech_prob": 3.882096033525073e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 518, "review_annotation": "", "row_id": 518, "bounded_row_id": 518, "parent_row_ids": [ 518 ], "parent_row_id": 518, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "这一操作将视觉tokens的数量", "selected_text": "这一操作将视觉tokens的数量", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 518 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1108.5594230769232, "end": 1110.0794230769231, "text": "压缩至原来的九分之一", "raw_asr_text": "压缩至原来的九分之一", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "压", "start": 1108.5594230769232, "end": 1108.7794230769232, "probability": 0.9990234375 }, { "word": "缩", "start": 1108.7794230769232, "end": 1108.9194230769233, "probability": 0.9990234375 }, { "word": "至", "start": 1108.9194230769233, "end": 1109.0394230769232, "probability": 0.9912109375 }, { "word": "原", "start": 1109.0394230769232, "end": 1109.1794230769233, "probability": 0.98486328125 }, { "word": "来", "start": 1109.1794230769233, "end": 1109.3794230769233, "probability": 0.99853515625 }, { "word": "的", "start": 1109.3794230769233, "end": 1109.4594230769233, "probability": 0.99169921875 }, { "word": "九", "start": 1109.4594230769233, "end": 1109.5994230769234, "probability": 0.55615234375 }, { "word": "分", "start": 1109.5994230769234, "end": 1109.7594230769232, "probability": 0.99755859375 }, { "word": "之", "start": 1109.7594230769232, "end": 1109.9194230769233, "probability": 0.9990234375 }, { "word": "一", "start": 1109.9194230769233, "end": 1110.0794230769231, "probability": 0.9931640625 } ], "asr_metrics": { "avg_logprob": -0.1282342020670573, "compression_ratio": 1.3015873015873016, "no_speech_prob": 3.882096033525073e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 519, "review_annotation": "", "row_id": 519, "bounded_row_id": 519, "parent_row_ids": [ 519 ], "parent_row_id": 519, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "压缩至原来的九分之一", "selected_text": "压缩至原来的九分之一", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 519 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1110.0794230769231, "end": 1111.4594230769233, "text": "同时仍支持分辨率", "raw_asr_text": "同时仍支持分辨率", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "同", "start": 1110.0794230769231, "end": 1110.3194230769232, "probability": 0.99755859375 }, { "word": "时", "start": 1110.3194230769232, "end": 1110.4994230769232, "probability": 0.99267578125 }, { "word": "仍", "start": 1110.4994230769232, "end": 1110.6794230769233, "probability": 0.9951171875 }, { "word": "支持", "start": 1110.6794230769233, "end": 1111.0194230769232, "probability": 0.955078125 }, { "word": "分", "start": 1111.0194230769232, "end": 1111.1794230769233, "probability": 0.99755859375 }, { "word": "辨", "start": 1111.1794230769233, "end": 1111.2994230769232, "probability": 0.99365234375 }, { "word": "率", "start": 1111.2994230769232, "end": 1111.4594230769233, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.1282342020670573, "compression_ratio": 1.3015873015873016, "no_speech_prob": 3.882096033525073e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 520, "review_annotation": "", "row_id": 520, "bounded_row_id": 520, "parent_row_ids": [ 520 ], "parent_row_id": 520, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "同时仍支持分辨率", "selected_text": "同时仍支持分辨率", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 520 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1111.4594230769233, "end": 1114.5794230769231, "text": "约为1344×1344的图像", "raw_asr_text": "约为1344×1344的图像", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "约", "start": 1111.4594230769233, "end": 1111.6194230769233, "probability": 0.98388671875 }, { "word": "为", "start": 1111.6194230769233, "end": 1111.7994230769232, "probability": 0.9970703125 }, { "word": "13", "start": 1111.7994230769232, "end": 1112.0994230769234, "probability": 0.970703125 }, { "word": "44", "start": 1112.0994230769234, "end": 1112.7794230769232, "probability": 0.998046875 }, { "word": "×", "start": 1112.7794230769232, "end": 1113.0994230769234, "probability": 0.3974609375 }, { "word": "13", "start": 1113.0994230769234, "end": 1113.3994230769233, "probability": 0.99755859375 }, { "word": "44", "start": 1113.3994230769233, "end": 1114.0394230769232, "probability": 0.99951171875 }, { "word": "的", "start": 1114.0394230769232, "end": 1114.2794230769232, "probability": 0.9697265625 }, { "word": "图", "start": 1114.2794230769232, "end": 1114.4394230769233, "probability": 0.9990234375 }, { "word": "像", "start": 1114.4394230769233, "end": 1114.5794230769231, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.1282342020670573, "compression_ratio": 1.3015873015873016, "no_speech_prob": 3.882096033525073e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 521, "review_annotation": "", "row_id": 521, "bounded_row_id": 521, "parent_row_ids": [ 521 ], "parent_row_id": 521, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "约为1344×1344的图像", "selected_text": "约为1344×1344的图像", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 521 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1114.5794230769231, "end": 1116.9994230769232, "text": "为了帮你理清我刚才讲到的所有内容", "raw_asr_text": "为了帮你理清我刚才讲到的所有内容", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "为", "start": 1114.5794230769231, "end": 1114.9194230769233, "probability": 0.97998046875 }, { "word": "了", "start": 1114.9194230769233, "end": 1114.9994230769232, "probability": 0.99951171875 }, { "word": "帮", "start": 1114.9994230769232, "end": 1115.0794230769231, "probability": 0.99951171875 }, { "word": "你", "start": 1115.0794230769231, "end": 1115.2594230769232, "probability": 0.9775390625 }, { "word": "理", "start": 1115.2594230769232, "end": 1115.3794230769233, "probability": 0.97802734375 }, { "word": "清", "start": 1115.3794230769233, "end": 1115.6394230769233, "probability": 0.97705078125 }, { "word": "我", "start": 1115.6394230769233, "end": 1115.8394230769231, "probability": 0.99365234375 }, { "word": "刚", "start": 1115.8394230769231, "end": 1115.9794230769232, "probability": 0.99853515625 }, { "word": "才", "start": 1115.9794230769232, "end": 1116.0994230769234, "probability": 0.9970703125 }, { "word": "讲", "start": 1116.0994230769234, "end": 1116.2394230769232, "probability": 0.99658203125 }, { "word": "到", "start": 1116.2394230769232, "end": 1116.3594230769233, "probability": 0.98828125 }, { "word": "的", "start": 1116.3594230769233, "end": 1116.4594230769233, "probability": 0.9951171875 }, { "word": "所有", "start": 1116.4594230769233, "end": 1116.6394230769233, "probability": 0.9970703125 }, { "word": "内", "start": 1116.6394230769233, "end": 1116.8194230769232, "probability": 0.99755859375 }, { "word": "容", "start": 1116.8194230769232, "end": 1116.9994230769232, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1282342020670573, "compression_ratio": 1.3015873015873016, "no_speech_prob": 3.882096033525073e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 522, "review_annotation": "", "row_id": 522, "bounded_row_id": 522, "parent_row_ids": [ 522 ], "parent_row_id": 522, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "为了帮你理清我刚才讲到的所有内容", "selected_text": "为了帮你理清我刚才讲到的所有内容", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 522 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1116.9994230769232, "end": 1118.1994230769233, "text": "我们来完整梳理一下", "raw_asr_text": "我们来完整梳理一下", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "我们", "start": 1116.9994230769232, "end": 1117.2594230769232, "probability": 0.99365234375 }, { "word": "来", "start": 1117.2594230769232, "end": 1117.3594230769233, "probability": 0.99755859375 }, { "word": "完", "start": 1117.3594230769233, "end": 1117.4994230769232, "probability": 0.9931640625 }, { "word": "整", "start": 1117.4994230769232, "end": 1117.6394230769233, "probability": 0.99951171875 }, { "word": "梳", "start": 1117.6394230769233, "end": 1117.7994230769232, "probability": 0.98388671875 }, { "word": "理", "start": 1117.7994230769232, "end": 1117.9394230769233, "probability": 1.0 }, { "word": "一下", "start": 1117.9394230769233, "end": 1118.1994230769233, "probability": 0.98828125 } ], "asr_metrics": { "avg_logprob": -0.1282342020670573, "compression_ratio": 1.3015873015873016, "no_speech_prob": 3.882096033525073e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 523, "review_annotation": "", "row_id": 523, "bounded_row_id": 523, "parent_row_ids": [ 523 ], "parent_row_id": 523, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "我们来完整梳理一下", "selected_text": "我们来完整梳理一下", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 523 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1118.1994230769233, "end": 1119.7994230769232, "text": "他们搭建的整套模型架构", "raw_asr_text": "他们搭建的整套模型架构", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "他们", "start": 1118.1994230769233, "end": 1118.4194230769233, "probability": 0.59716796875 }, { "word": "搭", "start": 1118.4194230769233, "end": 1118.5394230769232, "probability": 0.998046875 }, { "word": "建", "start": 1118.5394230769232, "end": 1118.7194230769232, "probability": 0.99951171875 }, { "word": "的", "start": 1118.7194230769232, "end": 1118.8394230769231, "probability": 0.99169921875 }, { "word": "整", "start": 1118.8394230769231, "end": 1119.0194230769232, "probability": 0.9970703125 }, { "word": "套", "start": 1119.0194230769232, "end": 1119.1994230769233, "probability": 1.0 }, { "word": "模", "start": 1119.1994230769233, "end": 1119.3394230769231, "probability": 0.99755859375 }, { "word": "型", "start": 1119.3394230769231, "end": 1119.4794230769232, "probability": 0.99951171875 }, { "word": "架", "start": 1119.4794230769232, "end": 1119.6394230769233, "probability": 0.9970703125 }, { "word": "构", "start": 1119.6394230769233, "end": 1119.7994230769232, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.1282342020670573, "compression_ratio": 1.3015873015873016, "no_speech_prob": 3.882096033525073e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 524, "review_annotation": "", "row_id": 524, "bounded_row_id": 524, "parent_row_ids": [ 524 ], "parent_row_id": 524, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "他们搭建的整套模型架构", "selected_text": "他们搭建的整套模型架构", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 524 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1119.7994230769232, "end": 1121.6394230769233, "text": "我们从最底层开始讲起", "raw_asr_text": "我们从最底层开始讲起", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "我们", "start": 1119.7994230769232, "end": 1120.1994230769233, "probability": 0.99365234375 }, { "word": "从", "start": 1120.1994230769233, "end": 1120.4194230769233, "probability": 0.99853515625 }, { "word": "最", "start": 1120.4194230769233, "end": 1120.5994230769234, "probability": 0.9638671875 }, { "word": "底", "start": 1120.5994230769234, "end": 1120.8594230769233, "probability": 0.99853515625 }, { "word": "层", "start": 1120.8594230769233, "end": 1121.0194230769232, "probability": 0.998046875 }, { "word": "开始", "start": 1121.0194230769232, "end": 1121.1994230769233, "probability": 0.99853515625 }, { "word": "讲", "start": 1121.1994230769233, "end": 1121.4394230769233, "probability": 0.998046875 }, { "word": "起", "start": 1121.4394230769233, "end": 1121.6394230769233, "probability": 0.99169921875 } ], "asr_metrics": { "avg_logprob": -0.1282342020670573, "compression_ratio": 1.3015873015873016, "no_speech_prob": 3.882096033525073e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 525, "review_annotation": "", "row_id": 525, "bounded_row_id": 525, "parent_row_ids": [ 525 ], "parent_row_id": 525, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "我们从最底层开始讲起", "selected_text": "我们从最底层开始讲起", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 525 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1121.6394230769233, "end": 1123.9194230769233, "text": "模型有文本和图像两类输入", "raw_asr_text": "模型有文本和图像两类输入", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "模", "start": 1121.6394230769233, "end": 1121.9794230769232, "probability": 0.98828125 }, { "word": "型", "start": 1121.9794230769232, "end": 1122.1394230769233, "probability": 1.0 }, { "word": "有", "start": 1122.1394230769233, "end": 1122.3194230769232, "probability": 0.994140625 }, { "word": "文", "start": 1122.3194230769232, "end": 1122.5794230769231, "probability": 0.98779296875 }, { "word": "本", "start": 1122.5794230769231, "end": 1122.7394230769232, "probability": 0.99462890625 }, { "word": "和", "start": 1122.7394230769232, "end": 1122.9594230769233, "probability": 0.99658203125 }, { "word": "图", "start": 1122.9594230769233, "end": 1123.1394230769233, "probability": 0.99951171875 }, { "word": "像", "start": 1123.1394230769233, "end": 1123.2594230769232, "probability": 0.99853515625 }, { "word": "两", "start": 1123.2594230769232, "end": 1123.4194230769233, "probability": 0.98193359375 }, { "word": "类", "start": 1123.4194230769233, "end": 1123.5794230769231, "probability": 0.9990234375 }, { "word": "输", "start": 1123.5794230769231, "end": 1123.7394230769232, "probability": 0.99951171875 }, { "word": "入", "start": 1123.7394230769232, "end": 1123.9194230769233, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.1282342020670573, "compression_ratio": 1.3015873015873016, "no_speech_prob": 3.882096033525073e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 526, "review_annotation": "", "row_id": 526, "bounded_row_id": 526, "parent_row_ids": [ 526 ], "parent_row_id": 526, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "模型有文本和图像两类输入", "selected_text": "模型有文本和图像两类输入", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 526 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1123.9194230769233, "end": 1124.7794230769232, "text": "二者输入模型后", "raw_asr_text": "二者输入模型后", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "二", "start": 1123.9194230769233, "end": 1124.0594230769232, "probability": 0.9931640625 }, { "word": "者", "start": 1124.0594230769232, "end": 1124.1794230769233, "probability": 0.98876953125 }, { "word": "输", "start": 1124.1794230769233, "end": 1124.2994230769232, "probability": 0.9990234375 }, { "word": "入", "start": 1124.2994230769232, "end": 1124.3994230769233, "probability": 1.0 }, { "word": "模", "start": 1124.3994230769233, "end": 1124.5194230769232, "probability": 0.99267578125 }, { "word": "型", "start": 1124.5194230769232, "end": 1124.6594230769233, "probability": 0.99951171875 }, { "word": "后", "start": 1124.6594230769233, "end": 1124.7794230769232, "probability": 0.9443359375 } ], "asr_metrics": { "avg_logprob": -0.1282342020670573, "compression_ratio": 1.3015873015873016, "no_speech_prob": 3.882096033525073e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 527, "review_annotation": "", "row_id": 527, "bounded_row_id": 527, "parent_row_ids": [ 527 ], "parent_row_id": 527, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "二者输入模型后", "selected_text": "二者输入模型后", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 527 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1124.7794230769232, "end": 1126.4994230769232, "text": "会被合并为一个嵌入项量", "raw_asr_text": "会被合并为一个嵌入项量", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "会", "start": 1124.7794230769232, "end": 1124.9594230769233, "probability": 0.99169921875 }, { "word": "被", "start": 1124.9594230769233, "end": 1125.0994230769234, "probability": 0.98974609375 }, { "word": "合", "start": 1125.0994230769234, "end": 1125.2394230769232, "probability": 0.99853515625 }, { "word": "并", "start": 1125.2394230769232, "end": 1125.3794230769233, "probability": 0.98486328125 }, { "word": "为", "start": 1125.3794230769233, "end": 1125.4994230769232, "probability": 0.99658203125 }, { "word": "一个", "start": 1125.4994230769232, "end": 1125.7194230769232, "probability": 0.9931640625 }, { "word": "嵌", "start": 1125.7194230769232, "end": 1126.0394230769232, "probability": 0.99267578125 }, { "word": "入", "start": 1126.0394230769232, "end": 1126.1994230769233, "probability": 0.99365234375 }, { "word": "项", "start": 1126.1994230769233, "end": 1126.3594230769233, "probability": 0.7939453125 }, { "word": "量", "start": 1126.3594230769233, "end": 1126.4994230769232, "probability": 0.548828125 } ], "asr_metrics": { "avg_logprob": -0.1282342020670573, "compression_ratio": 1.3015873015873016, "no_speech_prob": 3.882096033525073e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 528, "review_annotation": "", "row_id": 528, "bounded_row_id": 528, "parent_row_ids": [ 528 ], "parent_row_id": 528, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "会被合并为一个嵌入项量", "selected_text": "会被合并为一个嵌入项量", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 528 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1126.4994230769232, "end": 1127.9194230769233, "text": "随后它会经过两个", "raw_asr_text": "随后它会经过两个", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "随", "start": 1126.4994230769232, "end": 1126.8594230769233, "probability": 0.9951171875 }, { "word": "后", "start": 1126.8594230769233, "end": 1126.9594230769233, "probability": 0.99951171875 }, { "word": "它", "start": 1126.9594230769233, "end": 1127.0794230769231, "probability": 0.91845703125 }, { "word": "会", "start": 1127.0794230769231, "end": 1127.2394230769232, "probability": 0.99853515625 }, { "word": "经", "start": 1127.2394230769232, "end": 1127.4194230769233, "probability": 0.998046875 }, { "word": "过", "start": 1127.4194230769233, "end": 1127.5794230769231, "probability": 0.99951171875 }, { "word": "两", "start": 1127.5794230769231, "end": 1127.7594230769232, "probability": 0.990234375 }, { "word": "个", "start": 1127.7594230769232, "end": 1127.9194230769233, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.1282342020670573, "compression_ratio": 1.3015873015873016, "no_speech_prob": 3.882096033525073e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 529, "review_annotation": "", "row_id": 529, "bounded_row_id": 529, "parent_row_ids": [ 529 ], "parent_row_id": 529, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "随后它会经过两个", "selected_text": "随后它会经过两个", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 529 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1127.9194230769233, "end": 1129.3394230769231, "text": "纯滑动窗口注意力层", "raw_asr_text": "纯滑动窗口注意力层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "纯", "start": 1127.9194230769233, "end": 1128.1594230769233, "probability": 0.9912109375 }, { "word": "滑", "start": 1128.1594230769233, "end": 1128.3394230769231, "probability": 0.9853515625 }, { "word": "动", "start": 1128.3394230769231, "end": 1128.4594230769233, "probability": 0.9365234375 }, { "word": "窗", "start": 1128.4594230769233, "end": 1128.5994230769234, "probability": 0.998046875 }, { "word": "口", "start": 1128.5994230769234, "end": 1128.7794230769232, "probability": 0.9990234375 }, { "word": "注意", "start": 1128.7794230769232, "end": 1128.9994230769232, "probability": 0.982421875 }, { "word": "力", "start": 1128.9994230769232, "end": 1129.1394230769233, "probability": 0.9609375 }, { "word": "层", "start": 1129.1394230769233, "end": 1129.3394230769231, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.1282342020670573, "compression_ratio": 1.3015873015873016, "no_speech_prob": 3.882096033525073e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 530, "review_annotation": "", "row_id": 530, "bounded_row_id": 530, "parent_row_ids": [ 530 ], "parent_row_id": 530, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "纯滑动窗口注意力层", "selected_text": "纯滑动窗口注意力层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 530 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1129.3394230769231, "end": 1131.1074086538463, "text": "这两层会读取Ngram提供的记忆之后", "raw_asr_text": "这两层会读取Ngram提供的记忆之后", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 40, "words": [ { "word": "这", "start": 1129.3394230769231, "end": 1129.6594230769233, "probability": 0.99609375 }, { "word": "两", "start": 1129.6594230769233, "end": 1129.7594230769232, "probability": 0.99755859375 }, { "word": "层", "start": 1129.7594230769232, "end": 1129.9394230769233, "probability": 0.9990234375 }, { "word": "会", "start": 1129.9394230769233, "end": 1130.1194230769233, "probability": 0.9990234375 }, { "word": "读", "start": 1130.1194230769233, "end": 1130.2794230769232, "probability": 0.9970703125 }, { "word": "取", "start": 1130.2794230769232, "end": 1130.4394230769233, "probability": 1.0 }, { "word": "N", "start": 1130.4394230769233, "end": 1130.6594230769233, "probability": 0.58642578125 }, { "word": "gram", "start": 1130.6594230769233, "end": 1130.8794230769233, "probability": 0.97314453125 }, { "word": "提", "start": 1130.8794230769233, "end": 1131.1194230769233, "probability": 0.99853515625 }, { "word": "供", "start": 1131.1194230769233, "end": 1131.2594230769232, "probability": 1.0 }, { "word": "的", "start": 1131.2594230769232, "end": 1131.3594230769233, "probability": 1.0 }, { "word": "记", "start": 1131.3594230769233, "end": 1131.4994230769232, "probability": 0.9501953125 }, { "word": "忆", "start": 1131.4994230769232, "end": 1131.6394230769233, "probability": 0.99609375 }, { "word": "之", "start": 1131.6394230769233, "end": 1131.7594230769232, "probability": 0.98583984375 }, { "word": "后", "start": 1131.7594230769232, "end": 1131.8794230769233, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.11160330338911577, "compression_ratio": 0.8, "no_speech_prob": 6.096781499792536e-12 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在混洗的方式压缩视觉特征再将其输入语言模型这一操作将视觉tokens的数量压缩至原来的九分之一同时仍支持分辨率约为1344×1344的图像为了帮你理清我刚才讲到的所有内容我们来完整梳理一下他们搭建的整套模型架构我们从最底层开始讲起模型有文本和图像两类输入二者输入模型后会被合并为一个嵌入项量随后它会经过两个纯滑动窗口注意力层这两层会读取Ngram提供的记忆之后" } ], "id": 531, "review_annotation": "", "row_id": 531, "bounded_row_id": 531, "parent_row_ids": [ 531 ], "parent_row_id": 531, "packet_id": 40, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1102.7194230769232, "packet_end": 1131.9074086538462, "raw_text": "这两层会读取Ngram提供的记忆之后", "selected_text": "这两层会读取Ngram提供的记忆之后", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 531 ], "packet_id": 40, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1131.6274086538463, "end": 1135.1874086538464, "text": "之后会进入配比为1-5的CESAR", "raw_asr_text": "之后会进入配比为1-5的CESAR", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 41, "words": [ { "word": "之", "start": 1131.6274086538463, "end": 1131.7874086538463, "probability": 0.87060546875 }, { "word": "后", "start": 1131.7874086538463, "end": 1131.9274086538462, "probability": 1.0 }, { "word": "会", "start": 1131.9274086538462, "end": 1132.1474086538462, "probability": 0.9677734375 }, { "word": "进", "start": 1132.1474086538462, "end": 1132.3074086538463, "probability": 0.99462890625 }, { "word": "入", "start": 1132.3074086538463, "end": 1132.5274086538464, "probability": 0.99853515625 }, { "word": "配", "start": 1132.5274086538464, "end": 1132.7674086538464, "probability": 0.9765625 }, { "word": "比", "start": 1132.7674086538464, "end": 1132.9474086538464, "probability": 0.9541015625 }, { "word": "为", "start": 1132.9474086538464, "end": 1133.2074086538464, "probability": 0.97802734375 }, { "word": "1", "start": 1133.2074086538464, "end": 1133.4874086538464, "probability": 0.955078125 }, { "word": "-", "start": 1133.4874086538464, "end": 1133.6074086538463, "probability": 0.049163818359375 }, { "word": "5", "start": 1133.6074086538463, "end": 1133.8274086538463, "probability": 0.998046875 }, { "word": "的", "start": 1133.8274086538463, "end": 1133.9874086538464, "probability": 0.865234375 }, { "word": "C", "start": 1133.9874086538464, "end": 1134.3474086538463, "probability": 0.6533203125 }, { "word": "ES", "start": 1134.3474086538463, "end": 1134.7474086538464, "probability": 0.833984375 }, { "word": "AR", "start": 1134.7474086538464, "end": 1135.1874086538464, "probability": 0.69970703125 } ], "asr_metrics": { "avg_logprob": -0.17342325846354167, "compression_ratio": 1.1917808219178083, "no_speech_prob": 5.060477037410749e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "允许Ngram提供的记忆之后会进入配比为1-5的CESAR全量层与复用层该结构会重复三次这样加起来一共是18层算下来英国编码器组件的总层数为20层这里的数字2指的是大约每两个TOKENS会被压缩成一个主KV条目且每个CESAR层都配套有滑动窗口注意力至于MOE部分他们仍在使用DeepseekMOE到这一步为止的所有输出就是供后续层使用的编码器引状态好" } ], "id": 532, "review_annotation": "", "row_id": 532, "bounded_row_id": 532, "parent_row_ids": [ 532 ], "parent_row_id": 532, "packet_id": 41, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1130.3074086538463, "packet_end": 1159.4953942307693, "raw_text": "之后会进入配比为1-5的CESAR", "selected_text": "之后会进入配比为1-5的CESAR", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 532 ], "packet_id": 41, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1135.1874086538464, "end": 1136.5474086538463, "text": "全量层与复用层", "raw_asr_text": "全量层与复用层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 41, "words": [ { "word": "全", "start": 1135.1874086538464, "end": 1135.5474086538463, "probability": 0.876953125 }, { "word": "量", "start": 1135.5474086538463, "end": 1135.7674086538464, "probability": 0.974609375 }, { "word": "层", "start": 1135.7674086538464, "end": 1135.9674086538464, "probability": 0.9990234375 }, { "word": "与", "start": 1135.9674086538464, "end": 1136.1274086538463, "probability": 0.98876953125 }, { "word": "复", "start": 1136.1274086538463, "end": 1136.2674086538464, "probability": 0.798828125 }, { "word": "用", "start": 1136.2674086538464, "end": 1136.3674086538463, "probability": 0.99853515625 }, { "word": "层", "start": 1136.3674086538463, "end": 1136.5474086538463, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.17342325846354167, "compression_ratio": 1.1917808219178083, "no_speech_prob": 5.060477037410749e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "允许Ngram提供的记忆之后会进入配比为1-5的CESAR全量层与复用层该结构会重复三次这样加起来一共是18层算下来英国编码器组件的总层数为20层这里的数字2指的是大约每两个TOKENS会被压缩成一个主KV条目且每个CESAR层都配套有滑动窗口注意力至于MOE部分他们仍在使用DeepseekMOE到这一步为止的所有输出就是供后续层使用的编码器引状态好" } ], "id": 533, "review_annotation": "", "row_id": 533, "bounded_row_id": 533, "parent_row_ids": [ 533 ], "parent_row_id": 533, "packet_id": 41, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1130.3074086538463, "packet_end": 1159.4953942307693, "raw_text": "全量层与复用层", "selected_text": "全量层与复用层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 533 ], "packet_id": 41, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1136.5474086538463, "end": 1137.8874086538463, "text": "该结构会重复三次", "raw_asr_text": "该结构会重复三次", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 41, "words": [ { "word": "该", "start": 1136.5474086538463, "end": 1136.7474086538464, "probability": 0.91357421875 }, { "word": "结", "start": 1136.7474086538464, "end": 1136.9274086538462, "probability": 0.99755859375 }, { "word": "构", "start": 1136.9274086538462, "end": 1137.1074086538463, "probability": 0.9990234375 }, { "word": "会", "start": 1137.1074086538463, "end": 1137.2074086538464, "probability": 0.9736328125 }, { "word": "重", "start": 1137.2074086538464, "end": 1137.3474086538463, "probability": 0.9970703125 }, { "word": "复", "start": 1137.3474086538463, "end": 1137.5074086538464, "probability": 0.9912109375 }, { "word": "三", "start": 1137.5074086538464, "end": 1137.6674086538462, "probability": 0.615234375 }, { "word": "次", "start": 1137.6674086538462, "end": 1137.8874086538463, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.17342325846354167, "compression_ratio": 1.1917808219178083, "no_speech_prob": 5.060477037410749e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "允许Ngram提供的记忆之后会进入配比为1-5的CESAR全量层与复用层该结构会重复三次这样加起来一共是18层算下来英国编码器组件的总层数为20层这里的数字2指的是大约每两个TOKENS会被压缩成一个主KV条目且每个CESAR层都配套有滑动窗口注意力至于MOE部分他们仍在使用DeepseekMOE到这一步为止的所有输出就是供后续层使用的编码器引状态好" } ], "id": 534, "review_annotation": "", "row_id": 534, "bounded_row_id": 534, "parent_row_ids": [ 534 ], "parent_row_id": 534, "packet_id": 41, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1130.3074086538463, "packet_end": 1159.4953942307693, "raw_text": "该结构会重复三次", "selected_text": "该结构会重复三次", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 534 ], "packet_id": 41, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1137.8874086538463, "end": 1140.2274086538464, "text": "这样加起来一共是18层", "raw_asr_text": "这样加起来一共是18层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 41, "words": [ { "word": "这样", "start": 1137.8874086538463, "end": 1138.1474086538462, "probability": 0.9384765625 }, { "word": "加", "start": 1138.1474086538462, "end": 1138.3874086538463, "probability": 0.98974609375 }, { "word": "起", "start": 1138.3874086538463, "end": 1138.5674086538463, "probability": 0.9951171875 }, { "word": "来", "start": 1138.5674086538463, "end": 1138.7074086538464, "probability": 0.9970703125 }, { "word": "一", "start": 1138.7074086538464, "end": 1138.9274086538462, "probability": 0.96875 }, { "word": "共", "start": 1138.9274086538462, "end": 1139.0874086538463, "probability": 0.9990234375 }, { "word": "是", "start": 1139.0874086538463, "end": 1139.3074086538463, "probability": 0.90380859375 }, { "word": "18", "start": 1139.3074086538463, "end": 1139.8474086538463, "probability": 0.97998046875 }, { "word": "层", "start": 1139.8474086538463, "end": 1140.2274086538464, "probability": 0.994140625 } ], "asr_metrics": { "avg_logprob": -0.17342325846354167, "compression_ratio": 1.1917808219178083, "no_speech_prob": 5.060477037410749e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "允许Ngram提供的记忆之后会进入配比为1-5的CESAR全量层与复用层该结构会重复三次这样加起来一共是18层算下来英国编码器组件的总层数为20层这里的数字2指的是大约每两个TOKENS会被压缩成一个主KV条目且每个CESAR层都配套有滑动窗口注意力至于MOE部分他们仍在使用DeepseekMOE到这一步为止的所有输出就是供后续层使用的编码器引状态好" } ], "id": 535, "review_annotation": "", "row_id": 535, "bounded_row_id": 535, "parent_row_ids": [ 535 ], "parent_row_id": 535, "packet_id": 41, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1130.3074086538463, "packet_end": 1159.4953942307693, "raw_text": "这样加起来一共是18层", "selected_text": "这样加起来一共是18层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 535 ], "packet_id": 41, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1140.2274086538464, "end": 1143.3474086538463, "text": "算下来英国编码器组件的总层数为20层", "raw_asr_text": "算下来英国编码器组件的总层数为20层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 41, "words": [ { "word": "算", "start": 1140.2274086538464, "end": 1140.4874086538464, "probability": 0.9638671875 }, { "word": "下", "start": 1140.4874086538464, "end": 1140.6274086538463, "probability": 0.98388671875 }, { "word": "来", "start": 1140.6274086538463, "end": 1140.7674086538464, "probability": 0.958984375 }, { "word": "英", "start": 1140.7674086538464, "end": 1140.9474086538464, "probability": 0.873046875 }, { "word": "国", "start": 1140.9474086538464, "end": 1141.1074086538463, "probability": 0.9912109375 }, { "word": "编", "start": 1141.1074086538463, "end": 1141.3074086538463, "probability": 0.990234375 }, { "word": "码", "start": 1141.3074086538463, "end": 1141.4674086538464, "probability": 0.9970703125 }, { "word": "器", "start": 1141.4674086538464, "end": 1141.6274086538463, "probability": 0.99755859375 }, { "word": "组", "start": 1141.6274086538463, "end": 1141.8274086538463, "probability": 0.99755859375 }, { "word": "件", "start": 1141.8274086538463, "end": 1141.9674086538464, "probability": 0.9384765625 }, { "word": "的", "start": 1141.9674086538464, "end": 1142.1074086538463, "probability": 0.98193359375 }, { "word": "总", "start": 1142.1074086538463, "end": 1142.3074086538463, "probability": 0.9951171875 }, { "word": "层", "start": 1142.3074086538463, "end": 1142.4874086538464, "probability": 0.9814453125 }, { "word": "数", "start": 1142.4874086538464, "end": 1142.6474086538462, "probability": 0.9970703125 }, { "word": "为", "start": 1142.6474086538462, "end": 1142.8274086538463, "probability": 0.9921875 }, { "word": "20", "start": 1142.8274086538463, "end": 1143.0874086538463, "probability": 0.974609375 }, { "word": "层", "start": 1143.0874086538463, "end": 1143.3474086538463, "probability": 0.99365234375 } ], "asr_metrics": { "avg_logprob": -0.17342325846354167, "compression_ratio": 1.1917808219178083, "no_speech_prob": 5.060477037410749e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "允许Ngram提供的记忆之后会进入配比为1-5的CESAR全量层与复用层该结构会重复三次这样加起来一共是18层算下来英国编码器组件的总层数为20层这里的数字2指的是大约每两个TOKENS会被压缩成一个主KV条目且每个CESAR层都配套有滑动窗口注意力至于MOE部分他们仍在使用DeepseekMOE到这一步为止的所有输出就是供后续层使用的编码器引状态好" } ], "id": 536, "review_annotation": "", "row_id": 536, "bounded_row_id": 536, "parent_row_ids": [ 536 ], "parent_row_id": 536, "packet_id": 41, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1130.3074086538463, "packet_end": 1159.4953942307693, "raw_text": "算下来英国编码器组件的总层数为20层", "selected_text": "算下来英国编码器组件的总层数为20层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 536 ], "packet_id": 41, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1143.3474086538463, "end": 1144.7074086538464, "text": "这里的数字2指的是", "raw_asr_text": "这里的数字2指的是", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 41, "words": [ { "word": "这里", "start": 1143.3474086538463, "end": 1143.6674086538462, "probability": 0.98388671875 }, { "word": "的", "start": 1143.6674086538462, "end": 1143.8274086538463, "probability": 0.98828125 }, { "word": "数", "start": 1143.8274086538463, "end": 1143.9474086538464, "probability": 0.994140625 }, { "word": "字", "start": 1143.9474086538464, "end": 1144.1074086538463, "probability": 0.99755859375 }, { "word": "2", "start": 1144.1074086538463, "end": 1144.2474086538464, "probability": 0.72900390625 }, { "word": "指", "start": 1144.2474086538464, "end": 1144.4474086538464, "probability": 0.96484375 }, { "word": "的是", "start": 1144.4474086538464, "end": 1144.7074086538464, "probability": 0.990234375 } ], "asr_metrics": { "avg_logprob": -0.17342325846354167, "compression_ratio": 1.1917808219178083, "no_speech_prob": 5.060477037410749e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "允许Ngram提供的记忆之后会进入配比为1-5的CESAR全量层与复用层该结构会重复三次这样加起来一共是18层算下来英国编码器组件的总层数为20层这里的数字2指的是大约每两个TOKENS会被压缩成一个主KV条目且每个CESAR层都配套有滑动窗口注意力至于MOE部分他们仍在使用DeepseekMOE到这一步为止的所有输出就是供后续层使用的编码器引状态好" } ], "id": 537, "review_annotation": "", "row_id": 537, "bounded_row_id": 537, "parent_row_ids": [ 537 ], "parent_row_id": 537, "packet_id": 41, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1130.3074086538463, "packet_end": 1159.4953942307693, "raw_text": "这里的数字2指的是", "selected_text": "这里的数字2指的是", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 537 ], "packet_id": 41, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1144.7074086538464, "end": 1146.1674086538462, "text": "大约每两个TOKENS", "raw_asr_text": "大约每两个TOKENS", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 41, "words": [ { "word": "大", "start": 1144.7074086538464, "end": 1144.9274086538462, "probability": 0.98095703125 }, { "word": "约", "start": 1144.9274086538462, "end": 1145.1074086538463, "probability": 0.9990234375 }, { "word": "每", "start": 1145.1074086538463, "end": 1145.3274086538463, "probability": 0.99462890625 }, { "word": "两", "start": 1145.3274086538463, "end": 1145.5674086538463, "probability": 0.88427734375 }, { "word": "个", "start": 1145.5674086538463, "end": 1145.7474086538464, "probability": 0.99755859375 }, { "word": "T", "start": 1145.7474086538464, "end": 1145.8874086538463, "probability": 0.521484375 }, { "word": "OK", "start": 1145.8874086538463, "end": 1145.9874086538464, "probability": 0.4560546875 }, { "word": "ENS", "start": 1145.9874086538464, "end": 1146.1674086538462, "probability": 0.7666015625 } ], "asr_metrics": { "avg_logprob": -0.17342325846354167, "compression_ratio": 1.1917808219178083, "no_speech_prob": 5.060477037410749e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "允许Ngram提供的记忆之后会进入配比为1-5的CESAR全量层与复用层该结构会重复三次这样加起来一共是18层算下来英国编码器组件的总层数为20层这里的数字2指的是大约每两个TOKENS会被压缩成一个主KV条目且每个CESAR层都配套有滑动窗口注意力至于MOE部分他们仍在使用DeepseekMOE到这一步为止的所有输出就是供后续层使用的编码器引状态好" } ], "id": 538, "review_annotation": "", "row_id": 538, "bounded_row_id": 538, "parent_row_ids": [ 538 ], "parent_row_id": 538, "packet_id": 41, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1130.3074086538463, "packet_end": 1159.4953942307693, "raw_text": "大约每两个TOKENS", "selected_text": "大约每两个TOKENS", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 538 ], "packet_id": 41, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1146.1674086538462, "end": 1148.4274086538462, "text": "会被压缩成一个主KV条目", "raw_asr_text": "会被压缩成一个主KV条目", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 41, "words": [ { "word": "会", "start": 1146.1674086538462, "end": 1146.4074086538462, "probability": 0.9736328125 }, { "word": "被", "start": 1146.4074086538462, "end": 1146.5474086538463, "probability": 0.99951171875 }, { "word": "压", "start": 1146.5474086538463, "end": 1146.7674086538464, "probability": 0.9990234375 }, { "word": "缩", "start": 1146.7674086538464, "end": 1146.9474086538464, "probability": 0.9990234375 }, { "word": "成", "start": 1146.9474086538464, "end": 1147.1474086538462, "probability": 0.9931640625 }, { "word": "一个", "start": 1147.1474086538462, "end": 1147.5274086538464, "probability": 0.97119140625 }, { "word": "主", "start": 1147.5274086538464, "end": 1147.8074086538463, "probability": 0.97705078125 }, { "word": "K", "start": 1147.8074086538463, "end": 1147.9474086538464, "probability": 0.9716796875 }, { "word": "V", "start": 1147.9474086538464, "end": 1148.0874086538463, "probability": 0.99658203125 }, { "word": "条", "start": 1148.0874086538463, "end": 1148.2274086538464, "probability": 0.98779296875 }, { "word": "目", "start": 1148.2274086538464, "end": 1148.4274086538462, "probability": 0.958984375 } ], "asr_metrics": { "avg_logprob": -0.17342325846354167, "compression_ratio": 1.1917808219178083, "no_speech_prob": 5.060477037410749e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "允许Ngram提供的记忆之后会进入配比为1-5的CESAR全量层与复用层该结构会重复三次这样加起来一共是18层算下来英国编码器组件的总层数为20层这里的数字2指的是大约每两个TOKENS会被压缩成一个主KV条目且每个CESAR层都配套有滑动窗口注意力至于MOE部分他们仍在使用DeepseekMOE到这一步为止的所有输出就是供后续层使用的编码器引状态好" } ], "id": 539, "review_annotation": "", "row_id": 539, "bounded_row_id": 539, "parent_row_ids": [ 539 ], "parent_row_id": 539, "packet_id": 41, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1130.3074086538463, "packet_end": 1159.4953942307693, "raw_text": "会被压缩成一个主KV条目", "selected_text": "会被压缩成一个主KV条目", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 539 ], "packet_id": 41, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1148.4274086538462, "end": 1151.4674086538464, "text": "且每个CESAR层都配套有滑动窗口注意力", "raw_asr_text": "且每个CESAR层都配套有滑动窗口注意力", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 41, "words": [ { "word": "且", "start": 1148.4274086538462, "end": 1148.7674086538464, "probability": 0.92578125 }, { "word": "每", "start": 1148.7674086538464, "end": 1148.9474086538464, "probability": 0.99462890625 }, { "word": "个", "start": 1148.9474086538464, "end": 1149.1074086538463, "probability": 0.99462890625 }, { "word": "C", "start": 1149.1074086538463, "end": 1149.2274086538464, "probability": 0.79052734375 }, { "word": "ES", "start": 1149.2274086538464, "end": 1149.3674086538463, "probability": 0.98583984375 }, { "word": "AR", "start": 1149.3674086538463, "end": 1149.6274086538463, "probability": 0.99609375 }, { "word": "层", "start": 1149.6274086538463, "end": 1149.8474086538463, "probability": 0.9580078125 }, { "word": "都", "start": 1149.8474086538463, "end": 1150.0074086538464, "probability": 0.986328125 }, { "word": "配", "start": 1150.0074086538464, "end": 1150.1474086538462, "probability": 0.9990234375 }, { "word": "套", "start": 1150.1474086538462, "end": 1150.2674086538464, "probability": 0.9111328125 }, { "word": "有", "start": 1150.2674086538464, "end": 1150.4474086538464, "probability": 0.9873046875 }, { "word": "滑", "start": 1150.4474086538464, "end": 1150.7274086538464, "probability": 0.99072265625 }, { "word": "动", "start": 1150.7274086538464, "end": 1150.8474086538463, "probability": 0.99365234375 }, { "word": "窗", "start": 1150.8474086538463, "end": 1151.0074086538464, "probability": 0.99609375 }, { "word": "口", "start": 1151.0074086538464, "end": 1151.1274086538463, "probability": 0.9990234375 }, { "word": "注意", "start": 1151.1274086538463, "end": 1151.2874086538463, "probability": 0.97119140625 }, { "word": "力", "start": 1151.2874086538463, "end": 1151.4674086538464, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.17342325846354167, "compression_ratio": 1.1917808219178083, "no_speech_prob": 5.060477037410749e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "允许Ngram提供的记忆之后会进入配比为1-5的CESAR全量层与复用层该结构会重复三次这样加起来一共是18层算下来英国编码器组件的总层数为20层这里的数字2指的是大约每两个TOKENS会被压缩成一个主KV条目且每个CESAR层都配套有滑动窗口注意力至于MOE部分他们仍在使用DeepseekMOE到这一步为止的所有输出就是供后续层使用的编码器引状态好" } ], "id": 540, "review_annotation": "", "row_id": 540, "bounded_row_id": 540, "parent_row_ids": [ 540 ], "parent_row_id": 540, "packet_id": 41, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1130.3074086538463, "packet_end": 1159.4953942307693, "raw_text": "且每个CESAR层都配套有滑动窗口注意力", "selected_text": "且每个CESAR层都配套有滑动窗口注意力", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 540 ], "packet_id": 41, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1151.4674086538464, "end": 1152.5874086538463, "text": "至于MOE部分", "raw_asr_text": "至于MOE部分", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 41, "words": [ { "word": "至", "start": 1151.4674086538464, "end": 1151.9074086538462, "probability": 0.98779296875 }, { "word": "于", "start": 1151.9074086538462, "end": 1152.0074086538464, "probability": 0.9990234375 }, { "word": "MO", "start": 1152.0074086538464, "end": 1152.2474086538464, "probability": 0.82470703125 }, { "word": "E", "start": 1152.2474086538464, "end": 1152.3274086538463, "probability": 0.755859375 }, { "word": "部分", "start": 1152.3274086538463, "end": 1152.5874086538463, "probability": 0.939453125 } ], "asr_metrics": { "avg_logprob": -0.17342325846354167, "compression_ratio": 1.1917808219178083, "no_speech_prob": 5.060477037410749e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "允许Ngram提供的记忆之后会进入配比为1-5的CESAR全量层与复用层该结构会重复三次这样加起来一共是18层算下来英国编码器组件的总层数为20层这里的数字2指的是大约每两个TOKENS会被压缩成一个主KV条目且每个CESAR层都配套有滑动窗口注意力至于MOE部分他们仍在使用DeepseekMOE到这一步为止的所有输出就是供后续层使用的编码器引状态好" } ], "id": 541, "review_annotation": "", "row_id": 541, "bounded_row_id": 541, "parent_row_ids": [ 541 ], "parent_row_id": 541, "packet_id": 41, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1130.3074086538463, "packet_end": 1159.4953942307693, "raw_text": "至于MOE部分", "selected_text": "至于MOE部分", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 541 ], "packet_id": 41, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1152.5874086538463, "end": 1154.1874086538464, "text": "他们仍在使用DeepseekMOE", "raw_asr_text": "他们仍在使用DeepseekMOE", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 41, "words": [ { "word": "他们", "start": 1152.5874086538463, "end": 1152.9474086538464, "probability": 0.689453125 }, { "word": "仍", "start": 1152.9474086538464, "end": 1153.1074086538463, "probability": 0.9912109375 }, { "word": "在", "start": 1153.1074086538463, "end": 1153.2674086538464, "probability": 0.9814453125 }, { "word": "使", "start": 1153.2674086538464, "end": 1153.4274086538462, "probability": 0.98681640625 }, { "word": "用", "start": 1153.4274086538462, "end": 1153.5474086538463, "probability": 0.99951171875 }, { "word": "De", "start": 1153.5474086538463, "end": 1153.6674086538462, "probability": 0.84423828125 }, { "word": "ep", "start": 1153.6674086538462, "end": 1153.7474086538464, "probability": 0.84228515625 }, { "word": "se", "start": 1153.7474086538464, "end": 1153.8274086538463, "probability": 0.363525390625 }, { "word": "ek", "start": 1153.8274086538463, "end": 1153.9274086538462, "probability": 0.9189453125 }, { "word": "MO", "start": 1153.9274086538462, "end": 1154.0474086538463, "probability": 0.301025390625 }, { "word": "E", "start": 1154.0474086538463, "end": 1154.1874086538464, "probability": 0.94140625 } ], "asr_metrics": { "avg_logprob": -0.17342325846354167, "compression_ratio": 1.1917808219178083, "no_speech_prob": 5.060477037410749e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "允许Ngram提供的记忆之后会进入配比为1-5的CESAR全量层与复用层该结构会重复三次这样加起来一共是18层算下来英国编码器组件的总层数为20层这里的数字2指的是大约每两个TOKENS会被压缩成一个主KV条目且每个CESAR层都配套有滑动窗口注意力至于MOE部分他们仍在使用DeepseekMOE到这一步为止的所有输出就是供后续层使用的编码器引状态好" } ], "id": 542, "review_annotation": "", "row_id": 542, "bounded_row_id": 542, "parent_row_ids": [ 542 ], "parent_row_id": 542, "packet_id": 41, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1130.3074086538463, "packet_end": 1159.4953942307693, "raw_text": "他们仍在使用DeepseekMOE", "selected_text": "他们仍在使用DeepseekMOE", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 542 ], "packet_id": 41, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1154.1874086538464, "end": 1156.4474086538464, "text": "到这一步为止的所有输出", "raw_asr_text": "到这一步为止的所有输出", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 41, "words": [ { "word": "到", "start": 1154.1874086538464, "end": 1154.4874086538464, "probability": 0.95849609375 }, { "word": "这", "start": 1154.4874086538464, "end": 1154.7474086538464, "probability": 0.98486328125 }, { "word": "一", "start": 1154.7474086538464, "end": 1154.8874086538463, "probability": 0.99462890625 }, { "word": "步", "start": 1154.8874086538463, "end": 1155.0074086538464, "probability": 0.990234375 }, { "word": "为", "start": 1155.0074086538464, "end": 1155.2474086538464, "probability": 0.98828125 }, { "word": "止", "start": 1155.2474086538464, "end": 1155.4674086538464, "probability": 0.99853515625 }, { "word": "的", "start": 1155.4674086538464, "end": 1155.6074086538463, "probability": 0.96044921875 }, { "word": "所有", "start": 1155.6074086538463, "end": 1156.0074086538464, "probability": 0.9892578125 }, { "word": "输", "start": 1156.0074086538464, "end": 1156.2874086538463, "probability": 0.9990234375 }, { "word": "出", "start": 1156.2874086538463, "end": 1156.4474086538464, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.17342325846354167, "compression_ratio": 1.1917808219178083, "no_speech_prob": 5.060477037410749e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "允许Ngram提供的记忆之后会进入配比为1-5的CESAR全量层与复用层该结构会重复三次这样加起来一共是18层算下来英国编码器组件的总层数为20层这里的数字2指的是大约每两个TOKENS会被压缩成一个主KV条目且每个CESAR层都配套有滑动窗口注意力至于MOE部分他们仍在使用DeepseekMOE到这一步为止的所有输出就是供后续层使用的编码器引状态好" } ], "id": 543, "review_annotation": "", "row_id": 543, "bounded_row_id": 543, "parent_row_ids": [ 543 ], "parent_row_id": 543, "packet_id": 41, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1130.3074086538463, "packet_end": 1159.4953942307693, "raw_text": "到这一步为止的所有输出", "selected_text": "到这一步为止的所有输出", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 543 ], "packet_id": 41, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1156.4474086538464, "end": 1158.6953942307694, "text": "就是供后续层使用的编码器引状态", "raw_asr_text": "就是供后续层使用的编码器引状态", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 41, "words": [ { "word": "就是", "start": 1156.4474086538464, "end": 1156.7674086538464, "probability": 0.98486328125 }, { "word": "供", "start": 1156.7674086538464, "end": 1157.1274086538463, "probability": 0.86767578125 }, { "word": "后", "start": 1157.1274086538463, "end": 1157.3474086538463, "probability": 0.9892578125 }, { "word": "续", "start": 1157.3474086538463, "end": 1157.5074086538464, "probability": 0.9990234375 }, { "word": "层", "start": 1157.5074086538464, "end": 1157.6674086538462, "probability": 0.99169921875 }, { "word": "使", "start": 1157.6674086538462, "end": 1157.8674086538463, "probability": 0.9873046875 }, { "word": "用", "start": 1157.8674086538463, "end": 1158.0474086538463, "probability": 1.0 }, { "word": "的", "start": 1158.0474086538463, "end": 1158.1274086538463, "probability": 0.99560546875 }, { "word": "编", "start": 1158.1274086538463, "end": 1158.3074086538463, "probability": 0.99560546875 }, { "word": "码", "start": 1158.3074086538463, "end": 1158.4474086538464, "probability": 1.0 }, { "word": "器", "start": 1158.4474086538464, "end": 1158.5874086538463, "probability": 0.99951171875 }, { "word": "引", "start": 1158.5874086538463, "end": 1158.7474086538464, "probability": 0.826171875 }, { "word": "状", "start": 1158.7474086538464, "end": 1158.9674086538464, "probability": 0.96728515625 }, { "word": "态", "start": 1158.9674086538464, "end": 1159.1674086538462, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.17342325846354167, "compression_ratio": 1.1917808219178083, "no_speech_prob": 5.060477037410749e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "允许Ngram提供的记忆之后会进入配比为1-5的CESAR全量层与复用层该结构会重复三次这样加起来一共是18层算下来英国编码器组件的总层数为20层这里的数字2指的是大约每两个TOKENS会被压缩成一个主KV条目且每个CESAR层都配套有滑动窗口注意力至于MOE部分他们仍在使用DeepseekMOE到这一步为止的所有输出就是供后续层使用的编码器引状态好" } ], "id": 544, "review_annotation": "", "row_id": 544, "bounded_row_id": 544, "parent_row_ids": [ 544 ], "parent_row_id": 544, "packet_id": 41, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1130.3074086538463, "packet_end": 1159.4953942307693, "raw_text": "就是供后续层使用的编码器引状态", "selected_text": "就是供后续层使用的编码器引状态", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 544 ], "packet_id": 41, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1159.1553942307694, "end": 1161.2153942307693, "text": "而SWA有介重放机制", "raw_asr_text": "而SWA有介重放机制", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 42, "words": [ { "word": "而", "start": 1159.1553942307694, "end": 1159.4153942307694, "probability": 0.9375 }, { "word": "S", "start": 1159.4153942307694, "end": 1159.6153942307694, "probability": 0.744140625 }, { "word": "WA", "start": 1159.6153942307694, "end": 1160.0553942307695, "probability": 0.98095703125 }, { "word": "有", "start": 1160.0553942307695, "end": 1160.3553942307694, "probability": 0.95556640625 }, { "word": "介", "start": 1160.3553942307694, "end": 1160.4953942307693, "probability": 0.59228515625 }, { "word": "重", "start": 1160.4953942307693, "end": 1160.7153942307693, "probability": 0.97216796875 }, { "word": "放", "start": 1160.7153942307693, "end": 1160.8953942307694, "probability": 0.9853515625 }, { "word": "机", "start": 1160.8953942307694, "end": 1161.0953942307694, "probability": 0.98388671875 }, { "word": "制", "start": 1161.0953942307694, "end": 1161.2153942307693, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.10244054158528645, "compression_ratio": 1.401840490797546, "no_speech_prob": 3.534807516136773e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "使用的编码器引状态而SWA有介重放机制会直接基于该引状态运行无需从解码器层依次获取数据因此预填充阶段只需要运行英国编码器组件而解码器运行时则需要两个组件同时启动对应的激活参数量为81解码器阶段则为160亿接下来讲解码器的具体层结构第一层是CSAR全量层之后是三层CSAR副用层接着是配比为1比3的重索引与副用结构该结构重复4次这里的数字已表示" } ], "id": 545, "review_annotation": "", "row_id": 545, "bounded_row_id": 545, "parent_row_ids": [ 545 ], "parent_row_id": 545, "packet_id": 42, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1157.8953942307694, "packet_end": 1187.0833798076924, "raw_text": "而SWA有介重放机制", "selected_text": "而SWA有介重放机制", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 545 ], "packet_id": 42, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1161.2153942307693, "end": 1163.0553942307695, "text": "会直接基于该引状态运行", "raw_asr_text": "会直接基于该引状态运行", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 42, "words": [ { "word": "会", "start": 1161.2153942307693, "end": 1161.4153942307694, "probability": 0.98095703125 }, { "word": "直接", "start": 1161.4153942307694, "end": 1161.7353942307693, "probability": 0.99853515625 }, { "word": "基", "start": 1161.7353942307693, "end": 1161.9153942307694, "probability": 0.75439453125 }, { "word": "于", "start": 1161.9153942307694, "end": 1162.0553942307695, "probability": 0.9951171875 }, { "word": "该", "start": 1162.0553942307695, "end": 1162.2353942307693, "probability": 0.99169921875 }, { "word": "引", "start": 1162.2353942307693, "end": 1162.3553942307694, "probability": 0.98974609375 }, { "word": "状", "start": 1162.3553942307694, "end": 1162.5353942307695, "probability": 1.0 }, { "word": "态", "start": 1162.5353942307695, "end": 1162.6753942307694, "probability": 1.0 }, { "word": "运", "start": 1162.6753942307694, "end": 1162.8553942307694, "probability": 0.9990234375 }, { "word": "行", "start": 1162.8553942307694, "end": 1163.0553942307695, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.10244054158528645, "compression_ratio": 1.401840490797546, "no_speech_prob": 3.534807516136773e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "使用的编码器引状态而SWA有介重放机制会直接基于该引状态运行无需从解码器层依次获取数据因此预填充阶段只需要运行英国编码器组件而解码器运行时则需要两个组件同时启动对应的激活参数量为81解码器阶段则为160亿接下来讲解码器的具体层结构第一层是CSAR全量层之后是三层CSAR副用层接着是配比为1比3的重索引与副用结构该结构重复4次这里的数字已表示" } ], "id": 546, "review_annotation": "", "row_id": 546, "bounded_row_id": 546, "parent_row_ids": [ 546 ], "parent_row_id": 546, "packet_id": 42, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1157.8953942307694, "packet_end": 1187.0833798076924, "raw_text": "会直接基于该引状态运行", "selected_text": "会直接基于该引状态运行", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 546 ], "packet_id": 42, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1163.0553942307695, "end": 1165.6953942307694, "text": "无需从解码器层依次获取数据", "raw_asr_text": "无需从解码器层依次获取数据", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 42, "words": [ { "word": "无", "start": 1163.0553942307695, "end": 1163.4753942307693, "probability": 0.896484375 }, { "word": "需", "start": 1163.4753942307693, "end": 1163.6353942307694, "probability": 0.9912109375 }, { "word": "从", "start": 1163.6353942307694, "end": 1163.8753942307694, "probability": 0.99365234375 }, { "word": "解", "start": 1163.8753942307694, "end": 1164.0353942307695, "probability": 0.873046875 }, { "word": "码", "start": 1164.0353942307695, "end": 1164.2153942307693, "probability": 1.0 }, { "word": "器", "start": 1164.2153942307693, "end": 1164.3753942307694, "probability": 0.99853515625 }, { "word": "层", "start": 1164.3753942307694, "end": 1164.5753942307695, "probability": 0.99462890625 }, { "word": "依", "start": 1164.5753942307695, "end": 1164.7953942307695, "probability": 0.970703125 }, { "word": "次", "start": 1164.7953942307695, "end": 1164.9753942307693, "probability": 0.99755859375 }, { "word": "获", "start": 1164.9753942307693, "end": 1165.1553942307694, "probability": 0.9990234375 }, { "word": "取", "start": 1165.1553942307694, "end": 1165.2953942307695, "probability": 0.99951171875 }, { "word": "数", "start": 1165.2953942307695, "end": 1165.4753942307693, "probability": 0.9970703125 }, { "word": "据", "start": 1165.4753942307693, "end": 1165.6953942307694, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10244054158528645, "compression_ratio": 1.401840490797546, "no_speech_prob": 3.534807516136773e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "使用的编码器引状态而SWA有介重放机制会直接基于该引状态运行无需从解码器层依次获取数据因此预填充阶段只需要运行英国编码器组件而解码器运行时则需要两个组件同时启动对应的激活参数量为81解码器阶段则为160亿接下来讲解码器的具体层结构第一层是CSAR全量层之后是三层CSAR副用层接着是配比为1比3的重索引与副用结构该结构重复4次这里的数字已表示" } ], "id": 547, "review_annotation": "", "row_id": 547, "bounded_row_id": 547, "parent_row_ids": [ 547 ], "parent_row_id": 547, "packet_id": 42, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1157.8953942307694, "packet_end": 1187.0833798076924, "raw_text": "无需从解码器层依次获取数据", "selected_text": "无需从解码器层依次获取数据", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 547 ], "packet_id": 42, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1165.6953942307694, "end": 1168.6953942307694, "text": "因此预填充阶段只需要运行英国编码器组件", "raw_asr_text": "因此预填充阶段只需要运行英国编码器组件", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 42, "words": [ { "word": "因", "start": 1165.6953942307694, "end": 1166.0153942307693, "probability": 0.98193359375 }, { "word": "此", "start": 1166.0153942307693, "end": 1166.1553942307694, "probability": 1.0 }, { "word": "预", "start": 1166.1553942307694, "end": 1166.2953942307695, "probability": 0.97021484375 }, { "word": "填", "start": 1166.2953942307695, "end": 1166.4353942307694, "probability": 0.9892578125 }, { "word": "充", "start": 1166.4353942307694, "end": 1166.5753942307695, "probability": 0.99609375 }, { "word": "阶", "start": 1166.5753942307695, "end": 1166.7353942307693, "probability": 0.9970703125 }, { "word": "段", "start": 1166.7353942307693, "end": 1166.8553942307694, "probability": 0.99951171875 }, { "word": "只", "start": 1166.8553942307694, "end": 1167.0353942307695, "probability": 0.9951171875 }, { "word": "需要", "start": 1167.0353942307695, "end": 1167.2953942307695, "probability": 0.982421875 }, { "word": "运", "start": 1167.2953942307695, "end": 1167.4553942307693, "probability": 0.99853515625 }, { "word": "行", "start": 1167.4553942307693, "end": 1167.6153942307694, "probability": 0.99853515625 }, { "word": "英", "start": 1167.6153942307694, "end": 1167.8353942307695, "probability": 0.8291015625 }, { "word": "国", "start": 1167.8353942307695, "end": 1167.9553942307693, "probability": 0.99462890625 }, { "word": "编", "start": 1167.9553942307693, "end": 1168.0953942307694, "probability": 0.99658203125 }, { "word": "码", "start": 1168.0953942307694, "end": 1168.2553942307693, "probability": 1.0 }, { "word": "器", "start": 1168.2553942307693, "end": 1168.3753942307694, "probability": 1.0 }, { "word": "组", "start": 1168.3753942307694, "end": 1168.5553942307695, "probability": 0.9970703125 }, { "word": "件", "start": 1168.5553942307695, "end": 1168.6953942307694, "probability": 0.99609375 } ], "asr_metrics": { "avg_logprob": -0.10244054158528645, "compression_ratio": 1.401840490797546, "no_speech_prob": 3.534807516136773e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "使用的编码器引状态而SWA有介重放机制会直接基于该引状态运行无需从解码器层依次获取数据因此预填充阶段只需要运行英国编码器组件而解码器运行时则需要两个组件同时启动对应的激活参数量为81解码器阶段则为160亿接下来讲解码器的具体层结构第一层是CSAR全量层之后是三层CSAR副用层接着是配比为1比3的重索引与副用结构该结构重复4次这里的数字已表示" } ], "id": 548, "review_annotation": "", "row_id": 548, "bounded_row_id": 548, "parent_row_ids": [ 548 ], "parent_row_id": 548, "packet_id": 42, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1157.8953942307694, "packet_end": 1187.0833798076924, "raw_text": "因此预填充阶段只需要运行英国编码器组件", "selected_text": "因此预填充阶段只需要运行英国编码器组件", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 548 ], "packet_id": 42, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1168.6953942307694, "end": 1171.6753942307694, "text": "而解码器运行时则需要两个组件同时启动", "raw_asr_text": "而解码器运行时则需要两个组件同时启动", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 42, "words": [ { "word": "而", "start": 1168.6953942307694, "end": 1169.0353942307695, "probability": 0.9833984375 }, { "word": "解", "start": 1169.0353942307695, "end": 1169.2353942307693, "probability": 0.99853515625 }, { "word": "码", "start": 1169.2353942307693, "end": 1169.3953942307694, "probability": 1.0 }, { "word": "器", "start": 1169.3953942307694, "end": 1169.5153942307693, "probability": 0.99951171875 }, { "word": "运", "start": 1169.5153942307693, "end": 1169.6553942307694, "probability": 0.99951171875 }, { "word": "行", "start": 1169.6553942307694, "end": 1169.7753942307695, "probability": 1.0 }, { "word": "时", "start": 1169.7753942307695, "end": 1169.9753942307693, "probability": 0.99560546875 }, { "word": "则", "start": 1169.9753942307693, "end": 1170.2153942307693, "probability": 0.99169921875 }, { "word": "需要", "start": 1170.2153942307693, "end": 1170.4553942307693, "probability": 0.99462890625 }, { "word": "两", "start": 1170.4553942307693, "end": 1170.6553942307694, "probability": 0.92724609375 }, { "word": "个", "start": 1170.6553942307694, "end": 1170.8353942307695, "probability": 0.99853515625 }, { "word": "组", "start": 1170.8353942307695, "end": 1170.9753942307693, "probability": 0.99951171875 }, { "word": "件", "start": 1170.9753942307693, "end": 1171.1153942307694, "probability": 1.0 }, { "word": "同", "start": 1171.1153942307694, "end": 1171.2553942307693, "probability": 0.9990234375 }, { "word": "时", "start": 1171.2553942307693, "end": 1171.3953942307694, "probability": 0.998046875 }, { "word": "启", "start": 1171.3953942307694, "end": 1171.5353942307695, "probability": 0.998046875 }, { "word": "动", "start": 1171.5353942307695, "end": 1171.6753942307694, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.10244054158528645, "compression_ratio": 1.401840490797546, "no_speech_prob": 3.534807516136773e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "使用的编码器引状态而SWA有介重放机制会直接基于该引状态运行无需从解码器层依次获取数据因此预填充阶段只需要运行英国编码器组件而解码器运行时则需要两个组件同时启动对应的激活参数量为81解码器阶段则为160亿接下来讲解码器的具体层结构第一层是CSAR全量层之后是三层CSAR副用层接着是配比为1比3的重索引与副用结构该结构重复4次这里的数字已表示" } ], "id": 549, "review_annotation": "", "row_id": 549, "bounded_row_id": 549, "parent_row_ids": [ 549 ], "parent_row_id": 549, "packet_id": 42, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1157.8953942307694, "packet_end": 1187.0833798076924, "raw_text": "而解码器运行时则需要两个组件同时启动", "selected_text": "而解码器运行时则需要两个组件同时启动", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 549 ], "packet_id": 42, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1171.6753942307694, "end": 1173.5553942307695, "text": "对应的激活参数量为81", "raw_asr_text": "对应的激活参数量为81", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 42, "words": [ { "word": "对", "start": 1171.6753942307694, "end": 1172.0753942307695, "probability": 0.9912109375 }, { "word": "应", "start": 1172.0753942307695, "end": 1172.1953942307694, "probability": 0.9990234375 }, { "word": "的", "start": 1172.1953942307694, "end": 1172.3553942307694, "probability": 0.9931640625 }, { "word": "激", "start": 1172.3553942307694, "end": 1172.4753942307693, "probability": 0.99658203125 }, { "word": "活", "start": 1172.4753942307693, "end": 1172.5953942307694, "probability": 0.9990234375 }, { "word": "参", "start": 1172.5953942307694, "end": 1172.7353942307693, "probability": 0.984375 }, { "word": "数", "start": 1172.7353942307693, "end": 1172.8553942307694, "probability": 0.99951171875 }, { "word": "量", "start": 1172.8553942307694, "end": 1173.0153942307693, "probability": 0.99853515625 }, { "word": "为", "start": 1173.0153942307693, "end": 1173.1753942307694, "probability": 0.99169921875 }, { "word": "81", "start": 1173.1753942307694, "end": 1173.5553942307695, "probability": 0.3427734375 } ], "asr_metrics": { "avg_logprob": -0.10244054158528645, "compression_ratio": 1.401840490797546, "no_speech_prob": 3.534807516136773e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "使用的编码器引状态而SWA有介重放机制会直接基于该引状态运行无需从解码器层依次获取数据因此预填充阶段只需要运行英国编码器组件而解码器运行时则需要两个组件同时启动对应的激活参数量为81解码器阶段则为160亿接下来讲解码器的具体层结构第一层是CSAR全量层之后是三层CSAR副用层接着是配比为1比3的重索引与副用结构该结构重复4次这里的数字已表示" } ], "id": 550, "review_annotation": "", "row_id": 550, "bounded_row_id": 550, "parent_row_ids": [ 550 ], "parent_row_id": 550, "packet_id": 42, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1157.8953942307694, "packet_end": 1187.0833798076924, "raw_text": "对应的激活参数量为81", "selected_text": "对应的激活参数量为81", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 550 ], "packet_id": 42, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1173.5553942307695, "end": 1175.3753942307694, "text": "解码器阶段则为160亿", "raw_asr_text": "解码器阶段则为160亿", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 42, "words": [ { "word": "解", "start": 1173.5553942307695, "end": 1173.9753942307693, "probability": 0.89697265625 }, { "word": "码", "start": 1173.9753942307693, "end": 1174.1153942307694, "probability": 1.0 }, { "word": "器", "start": 1174.1153942307694, "end": 1174.2353942307693, "probability": 1.0 }, { "word": "阶", "start": 1174.2353942307693, "end": 1174.3753942307694, "probability": 0.9970703125 }, { "word": "段", "start": 1174.3753942307694, "end": 1174.5153942307693, "probability": 1.0 }, { "word": "则", "start": 1174.5153942307693, "end": 1174.6753942307694, "probability": 0.9970703125 }, { "word": "为", "start": 1174.6753942307694, "end": 1174.7953942307695, "probability": 0.99853515625 }, { "word": "160", "start": 1174.7953942307695, "end": 1175.1353942307694, "probability": 0.98779296875 }, { "word": "亿", "start": 1175.1353942307694, "end": 1175.3753942307694, "probability": 0.9443359375 } ], "asr_metrics": { "avg_logprob": -0.10244054158528645, "compression_ratio": 1.401840490797546, "no_speech_prob": 3.534807516136773e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "使用的编码器引状态而SWA有介重放机制会直接基于该引状态运行无需从解码器层依次获取数据因此预填充阶段只需要运行英国编码器组件而解码器运行时则需要两个组件同时启动对应的激活参数量为81解码器阶段则为160亿接下来讲解码器的具体层结构第一层是CSAR全量层之后是三层CSAR副用层接着是配比为1比3的重索引与副用结构该结构重复4次这里的数字已表示" } ], "id": 551, "review_annotation": "", "row_id": 551, "bounded_row_id": 551, "parent_row_ids": [ 551 ], "parent_row_id": 551, "packet_id": 42, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1157.8953942307694, "packet_end": 1187.0833798076924, "raw_text": "解码器阶段则为160亿", "selected_text": "解码器阶段则为160亿", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 551 ], "packet_id": 42, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1175.3753942307694, "end": 1177.2353942307693, "text": "接下来讲解码器的具体层结构", "raw_asr_text": "接下来讲解码器的具体层结构", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 42, "words": [ { "word": "接", "start": 1175.3753942307694, "end": 1175.6353942307694, "probability": 0.99609375 }, { "word": "下", "start": 1175.6353942307694, "end": 1175.7153942307693, "probability": 0.99951171875 }, { "word": "来", "start": 1175.7153942307693, "end": 1175.7953942307695, "probability": 1.0 }, { "word": "讲", "start": 1175.7953942307695, "end": 1175.9353942307694, "probability": 0.99609375 }, { "word": "解", "start": 1175.9353942307694, "end": 1176.1153942307694, "probability": 0.9990234375 }, { "word": "码", "start": 1176.1153942307694, "end": 1176.2353942307693, "probability": 1.0 }, { "word": "器", "start": 1176.2353942307693, "end": 1176.3553942307694, "probability": 1.0 }, { "word": "的", "start": 1176.3553942307694, "end": 1176.4553942307693, "probability": 0.982421875 }, { "word": "具", "start": 1176.4553942307693, "end": 1176.6153942307694, "probability": 0.99755859375 }, { "word": "体", "start": 1176.6153942307694, "end": 1176.7753942307695, "probability": 1.0 }, { "word": "层", "start": 1176.7753942307695, "end": 1176.9353942307694, "probability": 0.99755859375 }, { "word": "结", "start": 1176.9353942307694, "end": 1177.0753942307695, "probability": 0.9912109375 }, { "word": "构", "start": 1177.0753942307695, "end": 1177.2353942307693, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10244054158528645, "compression_ratio": 1.401840490797546, "no_speech_prob": 3.534807516136773e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "使用的编码器引状态而SWA有介重放机制会直接基于该引状态运行无需从解码器层依次获取数据因此预填充阶段只需要运行英国编码器组件而解码器运行时则需要两个组件同时启动对应的激活参数量为81解码器阶段则为160亿接下来讲解码器的具体层结构第一层是CSAR全量层之后是三层CSAR副用层接着是配比为1比3的重索引与副用结构该结构重复4次这里的数字已表示" } ], "id": 552, "review_annotation": "", "row_id": 552, "bounded_row_id": 552, "parent_row_ids": [ 552 ], "parent_row_id": 552, "packet_id": 42, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1157.8953942307694, "packet_end": 1187.0833798076924, "raw_text": "接下来讲解码器的具体层结构", "selected_text": "接下来讲解码器的具体层结构", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 552 ], "packet_id": 42, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1177.2353942307693, "end": 1179.0953942307694, "text": "第一层是CSAR全量层", "raw_asr_text": "第一层是CSAR全量层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 42, "words": [ { "word": "第一", "start": 1177.2353942307693, "end": 1177.4553942307693, "probability": 0.89013671875 }, { "word": "层", "start": 1177.4553942307693, "end": 1177.6753942307694, "probability": 0.9990234375 }, { "word": "是", "start": 1177.6753942307694, "end": 1177.9353942307694, "probability": 0.99658203125 }, { "word": "CS", "start": 1177.9353942307694, "end": 1178.2753942307695, "probability": 0.63525390625 }, { "word": "AR", "start": 1178.2753942307695, "end": 1178.4753942307693, "probability": 0.92138671875 }, { "word": "全", "start": 1178.4753942307693, "end": 1178.7353942307693, "probability": 0.88916015625 }, { "word": "量", "start": 1178.7353942307693, "end": 1178.8753942307694, "probability": 0.99169921875 }, { "word": "层", "start": 1178.8753942307694, "end": 1179.0953942307694, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10244054158528645, "compression_ratio": 1.401840490797546, "no_speech_prob": 3.534807516136773e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "使用的编码器引状态而SWA有介重放机制会直接基于该引状态运行无需从解码器层依次获取数据因此预填充阶段只需要运行英国编码器组件而解码器运行时则需要两个组件同时启动对应的激活参数量为81解码器阶段则为160亿接下来讲解码器的具体层结构第一层是CSAR全量层之后是三层CSAR副用层接着是配比为1比3的重索引与副用结构该结构重复4次这里的数字已表示" } ], "id": 553, "review_annotation": "", "row_id": 553, "bounded_row_id": 553, "parent_row_ids": [ 553 ], "parent_row_id": 553, "packet_id": 42, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1157.8953942307694, "packet_end": 1187.0833798076924, "raw_text": "第一层是CSAR全量层", "selected_text": "第一层是CSAR全量层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 553 ], "packet_id": 42, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1179.0953942307694, "end": 1181.0353942307695, "text": "之后是三层CSAR副用层", "raw_asr_text": "之后是三层CSAR副用层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 42, "words": [ { "word": "之", "start": 1179.0953942307694, "end": 1179.2953942307695, "probability": 0.95361328125 }, { "word": "后", "start": 1179.2953942307695, "end": 1179.4553942307693, "probability": 1.0 }, { "word": "是", "start": 1179.4553942307693, "end": 1179.6553942307694, "probability": 0.9951171875 }, { "word": "三", "start": 1179.6553942307694, "end": 1179.8753942307694, "probability": 0.8701171875 }, { "word": "层", "start": 1179.8753942307694, "end": 1180.0753942307695, "probability": 1.0 }, { "word": "CS", "start": 1180.0753942307695, "end": 1180.2753942307695, "probability": 0.98388671875 }, { "word": "AR", "start": 1180.2753942307695, "end": 1180.4753942307693, "probability": 0.9990234375 }, { "word": "副", "start": 1180.4753942307693, "end": 1180.7153942307693, "probability": 0.61669921875 }, { "word": "用", "start": 1180.7153942307693, "end": 1180.8153942307695, "probability": 0.998046875 }, { "word": "层", "start": 1180.8153942307695, "end": 1181.0353942307695, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.10244054158528645, "compression_ratio": 1.401840490797546, "no_speech_prob": 3.534807516136773e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "使用的编码器引状态而SWA有介重放机制会直接基于该引状态运行无需从解码器层依次获取数据因此预填充阶段只需要运行英国编码器组件而解码器运行时则需要两个组件同时启动对应的激活参数量为81解码器阶段则为160亿接下来讲解码器的具体层结构第一层是CSAR全量层之后是三层CSAR副用层接着是配比为1比3的重索引与副用结构该结构重复4次这里的数字已表示" } ], "id": 554, "review_annotation": "", "row_id": 554, "bounded_row_id": 554, "parent_row_ids": [ 554 ], "parent_row_id": 554, "packet_id": 42, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1157.8953942307694, "packet_end": 1187.0833798076924, "raw_text": "之后是三层CSAR副用层", "selected_text": "之后是三层CSAR副用层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 554 ], "packet_id": 42, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1181.0353942307695, "end": 1184.2353942307693, "text": "接着是配比为1比3的重索引与副用结构", "raw_asr_text": "接着是配比为1比3的重索引与副用结构", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 42, "words": [ { "word": "接", "start": 1181.0353942307695, "end": 1181.2953942307695, "probability": 0.99560546875 }, { "word": "着", "start": 1181.2953942307695, "end": 1181.4553942307693, "probability": 0.99267578125 }, { "word": "是", "start": 1181.4553942307693, "end": 1181.6553942307694, "probability": 0.99658203125 }, { "word": "配", "start": 1181.6553942307694, "end": 1181.8753942307694, "probability": 0.99462890625 }, { "word": "比", "start": 1181.8753942307694, "end": 1182.0153942307693, "probability": 0.9912109375 }, { "word": "为", "start": 1182.0153942307693, "end": 1182.1953942307694, "probability": 0.99267578125 }, { "word": "1", "start": 1182.1953942307694, "end": 1182.4153942307694, "probability": 0.9423828125 }, { "word": "比", "start": 1182.4153942307694, "end": 1182.5353942307695, "probability": 0.01555633544921875 }, { "word": "3", "start": 1182.5353942307695, "end": 1182.7153942307693, "probability": 0.99658203125 }, { "word": "的", "start": 1182.7153942307693, "end": 1182.8353942307695, "probability": 0.97802734375 }, { "word": "重", "start": 1182.8353942307695, "end": 1183.0353942307695, "probability": 0.9990234375 }, { "word": "索", "start": 1183.0353942307695, "end": 1183.1953942307694, "probability": 0.8603515625 }, { "word": "引", "start": 1183.1953942307694, "end": 1183.3553942307694, "probability": 0.994140625 }, { "word": "与", "start": 1183.3553942307694, "end": 1183.6353942307694, "probability": 0.9931640625 }, { "word": "副", "start": 1183.6353942307694, "end": 1183.7553942307693, "probability": 0.99365234375 }, { "word": "用", "start": 1183.7553942307693, "end": 1183.8953942307694, "probability": 0.99951171875 }, { "word": "结", "start": 1183.8953942307694, "end": 1184.0353942307695, "probability": 0.99462890625 }, { "word": "构", "start": 1184.0353942307695, "end": 1184.2353942307693, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10244054158528645, "compression_ratio": 1.401840490797546, "no_speech_prob": 3.534807516136773e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "使用的编码器引状态而SWA有介重放机制会直接基于该引状态运行无需从解码器层依次获取数据因此预填充阶段只需要运行英国编码器组件而解码器运行时则需要两个组件同时启动对应的激活参数量为81解码器阶段则为160亿接下来讲解码器的具体层结构第一层是CSAR全量层之后是三层CSAR副用层接着是配比为1比3的重索引与副用结构该结构重复4次这里的数字已表示" } ], "id": 555, "review_annotation": "", "row_id": 555, "bounded_row_id": 555, "parent_row_ids": [ 555 ], "parent_row_id": 555, "packet_id": 42, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1157.8953942307694, "packet_end": 1187.0833798076924, "raw_text": "接着是配比为1比3的重索引与副用结构", "selected_text": "接着是配比为1比3的重索引与副用结构", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 555 ], "packet_id": 42, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1184.2353942307693, "end": 1185.8153942307695, "text": "该结构重复4次", "raw_asr_text": "该结构重复4次", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 42, "words": [ { "word": "该", "start": 1184.2353942307693, "end": 1184.5753942307695, "probability": 0.9833984375 }, { "word": "结", "start": 1184.5753942307695, "end": 1184.7753942307695, "probability": 0.998046875 }, { "word": "构", "start": 1184.7753942307695, "end": 1184.9353942307694, "probability": 1.0 }, { "word": "重", "start": 1184.9353942307694, "end": 1185.0953942307694, "probability": 0.99609375 }, { "word": "复", "start": 1185.0953942307694, "end": 1185.3553942307694, "probability": 0.994140625 }, { "word": "4", "start": 1185.3553942307694, "end": 1185.5953942307694, "probability": 0.7451171875 }, { "word": "次", "start": 1185.5953942307694, "end": 1185.8153942307695, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.10244054158528645, "compression_ratio": 1.401840490797546, "no_speech_prob": 3.534807516136773e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "使用的编码器引状态而SWA有介重放机制会直接基于该引状态运行无需从解码器层依次获取数据因此预填充阶段只需要运行英国编码器组件而解码器运行时则需要两个组件同时启动对应的激活参数量为81解码器阶段则为160亿接下来讲解码器的具体层结构第一层是CSAR全量层之后是三层CSAR副用层接着是配比为1比3的重索引与副用结构该结构重复4次这里的数字已表示" } ], "id": 556, "review_annotation": "", "row_id": 556, "bounded_row_id": 556, "parent_row_ids": [ 556 ], "parent_row_id": 556, "packet_id": 42, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1157.8953942307694, "packet_end": 1187.0833798076924, "raw_text": "该结构重复4次", "selected_text": "该结构重复4次", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 556 ], "packet_id": 42, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1186.2833798076924, "end": 1189.1833798076925, "text": "这里的数字已表示主KV不会进行Token维度的压缩", "raw_asr_text": "这里的数字已表示主KV不会进行Token维度的压缩", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 43, "words": [ { "word": "这里", "start": 1185.4833798076925, "end": 1186.0833798076924, "probability": 0.890625 }, { "word": "的", "start": 1186.0833798076924, "end": 1186.2433798076925, "probability": 0.98681640625 }, { "word": "数", "start": 1186.2433798076925, "end": 1186.3833798076926, "probability": 0.99755859375 }, { "word": "字", "start": 1186.3833798076926, "end": 1186.5633798076924, "probability": 0.99560546875 }, { "word": "已", "start": 1186.5633798076924, "end": 1186.6833798076925, "probability": 0.62060546875 }, { "word": "表示", "start": 1186.6833798076925, "end": 1186.8833798076926, "probability": 0.99462890625 }, { "word": "主", "start": 1186.8833798076926, "end": 1187.1833798076925, "probability": 0.9609375 }, { "word": "K", "start": 1187.1833798076925, "end": 1187.3633798076926, "probability": 0.8671875 }, { "word": "V", "start": 1187.3633798076926, "end": 1187.5233798076924, "probability": 0.9931640625 }, { "word": "不会", "start": 1187.5233798076924, "end": 1187.8233798076924, "probability": 0.97998046875 }, { "word": "进", "start": 1187.8233798076924, "end": 1188.0033798076925, "probability": 0.99755859375 }, { "word": "行", "start": 1188.0033798076925, "end": 1188.1633798076925, "probability": 0.9990234375 }, { "word": "T", "start": 1188.1633798076925, "end": 1188.3033798076924, "probability": 0.62890625 }, { "word": "oken", "start": 1188.3033798076924, "end": 1188.4833798076925, "probability": 0.98095703125 }, { "word": "维", "start": 1188.4833798076925, "end": 1188.6033798076924, "probability": 0.9619140625 }, { "word": "度", "start": 1188.6033798076924, "end": 1188.7433798076925, "probability": 0.99658203125 }, { "word": "的", "start": 1188.7433798076925, "end": 1188.8433798076924, "probability": 0.98388671875 }, { "word": "压", "start": 1188.8433798076924, "end": 1189.0233798076924, "probability": 0.9990234375 }, { "word": "缩", "start": 1189.0233798076924, "end": 1189.1833798076925, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.08884044880038695, "compression_ratio": 1.2506666666666666, "no_speech_prob": 3.118513780542287e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这里的数字已表示主KV不会进行Token维度的压缩我个人的理解是这大概率是为了保留完整的Token级分辨率毕竟解码器已经在跨层大规模复用编码器生成的全局KV了但即便采用了CSAR附用重索引层仍然需要重新扫描整个百万级Token的上下文因此DeepSeek加入了分层吸收索引器避免这类搜索再次产生过高的算力开销首个完整层会扫描全部内容构建一个最多包含约1.6万" } ], "id": 557, "review_annotation": "", "row_id": 557, "bounded_row_id": 557, "parent_row_ids": [ 557 ], "parent_row_id": 557, "packet_id": 43, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1185.4833798076925, "packet_end": 1214.6713653846155, "raw_text": "这里的数字已表示主KV不会进行Token维度的压缩", "selected_text": "这里的数字已表示主KV不会进行Token维度的压缩", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 557 ], "packet_id": 43, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1189.1833798076925, "end": 1190.6033798076924, "text": "我个人的理解是", "raw_asr_text": "我个人的理解是", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 43, "words": [ { "word": "我", "start": 1189.1833798076925, "end": 1189.4633798076925, "probability": 0.9248046875 }, { "word": "个", "start": 1189.4633798076925, "end": 1189.7033798076925, "probability": 0.998046875 }, { "word": "人", "start": 1189.7033798076925, "end": 1189.8833798076926, "probability": 1.0 }, { "word": "的", "start": 1189.8833798076926, "end": 1190.0033798076925, "probability": 0.986328125 }, { "word": "理", "start": 1190.0033798076925, "end": 1190.1633798076925, "probability": 0.99853515625 }, { "word": "解", "start": 1190.1633798076925, "end": 1190.3233798076924, "probability": 1.0 }, { "word": "是", "start": 1190.3233798076924, "end": 1190.6033798076924, "probability": 0.98974609375 } ], "asr_metrics": { "avg_logprob": -0.08884044880038695, "compression_ratio": 1.2506666666666666, "no_speech_prob": 3.118513780542287e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这里的数字已表示主KV不会进行Token维度的压缩我个人的理解是这大概率是为了保留完整的Token级分辨率毕竟解码器已经在跨层大规模复用编码器生成的全局KV了但即便采用了CSAR附用重索引层仍然需要重新扫描整个百万级Token的上下文因此DeepSeek加入了分层吸收索引器避免这类搜索再次产生过高的算力开销首个完整层会扫描全部内容构建一个最多包含约1.6万" } ], "id": 558, "review_annotation": "", "row_id": 558, "bounded_row_id": 558, "parent_row_ids": [ 558 ], "parent_row_id": 558, "packet_id": 43, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1185.4833798076925, "packet_end": 1214.6713653846155, "raw_text": "我个人的理解是", "selected_text": "我个人的理解是", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 558 ], "packet_id": 43, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1190.6033798076924, "end": 1194.4833798076925, "text": "这大概率是为了保留完整的Token级分辨率", "raw_asr_text": "这大概率是为了保留完整的Token级分辨率", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 43, "words": [ { "word": "这", "start": 1190.6033798076924, "end": 1191.0033798076925, "probability": 0.8212890625 }, { "word": "大概", "start": 1191.0033798076925, "end": 1191.3833798076926, "probability": 0.9990234375 }, { "word": "率", "start": 1191.3833798076926, "end": 1191.6233798076926, "probability": 0.99462890625 }, { "word": "是", "start": 1191.6233798076926, "end": 1191.8433798076924, "probability": 0.99609375 }, { "word": "为", "start": 1191.8433798076924, "end": 1192.0433798076924, "probability": 0.998046875 }, { "word": "了", "start": 1192.0433798076924, "end": 1192.2433798076925, "probability": 0.99951171875 }, { "word": "保", "start": 1192.2433798076925, "end": 1192.4033798076925, "probability": 0.9990234375 }, { "word": "留", "start": 1192.4033798076925, "end": 1192.6233798076926, "probability": 0.998046875 }, { "word": "完", "start": 1192.6233798076926, "end": 1192.8833798076926, "probability": 0.99658203125 }, { "word": "整", "start": 1192.8833798076926, "end": 1193.2433798076925, "probability": 0.99951171875 }, { "word": "的", "start": 1193.2433798076925, "end": 1193.4433798076925, "probability": 0.9951171875 }, { "word": "T", "start": 1193.4433798076925, "end": 1193.6433798076926, "probability": 0.90576171875 }, { "word": "oken", "start": 1193.6433798076926, "end": 1193.8233798076924, "probability": 0.9990234375 }, { "word": "级", "start": 1193.8233798076924, "end": 1194.0033798076925, "probability": 0.9140625 }, { "word": "分", "start": 1194.0033798076925, "end": 1194.1633798076925, "probability": 0.9951171875 }, { "word": "辨", "start": 1194.1633798076925, "end": 1194.3033798076924, "probability": 0.9892578125 }, { "word": "率", "start": 1194.3033798076924, "end": 1194.4833798076925, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.08884044880038695, "compression_ratio": 1.2506666666666666, "no_speech_prob": 3.118513780542287e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这里的数字已表示主KV不会进行Token维度的压缩我个人的理解是这大概率是为了保留完整的Token级分辨率毕竟解码器已经在跨层大规模复用编码器生成的全局KV了但即便采用了CSAR附用重索引层仍然需要重新扫描整个百万级Token的上下文因此DeepSeek加入了分层吸收索引器避免这类搜索再次产生过高的算力开销首个完整层会扫描全部内容构建一个最多包含约1.6万" } ], "id": 559, "review_annotation": "", "row_id": 559, "bounded_row_id": 559, "parent_row_ids": [ 559 ], "parent_row_id": 559, "packet_id": 43, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1185.4833798076925, "packet_end": 1214.6713653846155, "raw_text": "这大概率是为了保留完整的Token级分辨率", "selected_text": "这大概率是为了保留完整的Token级分辨率", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 559 ], "packet_id": 43, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1194.4833798076925, "end": 1199.3633798076926, "text": "毕竟解码器已经在跨层大规模复用编码器生成的全局KV了", "raw_asr_text": "毕竟解码器已经在跨层大规模复用编码器生成的全局KV了", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 43, "words": [ { "word": "毕", "start": 1194.4833798076925, "end": 1194.8833798076926, "probability": 0.98828125 }, { "word": "竟", "start": 1194.8833798076926, "end": 1195.0633798076924, "probability": 1.0 }, { "word": "解", "start": 1195.0633798076924, "end": 1195.2233798076925, "probability": 0.97998046875 }, { "word": "码", "start": 1195.2233798076925, "end": 1195.4033798076925, "probability": 0.998046875 }, { "word": "器", "start": 1195.4033798076925, "end": 1195.5433798076924, "probability": 0.9990234375 }, { "word": "已经", "start": 1195.5433798076924, "end": 1195.9033798076925, "probability": 0.994140625 }, { "word": "在", "start": 1195.9033798076925, "end": 1196.1433798076926, "probability": 0.99853515625 }, { "word": "跨", "start": 1196.1433798076926, "end": 1196.3633798076926, "probability": 0.99560546875 }, { "word": "层", "start": 1196.3633798076926, "end": 1196.5433798076924, "probability": 0.9970703125 }, { "word": "大", "start": 1196.5433798076924, "end": 1196.6833798076925, "probability": 0.99658203125 }, { "word": "规", "start": 1196.6833798076925, "end": 1196.8833798076926, "probability": 1.0 }, { "word": "模", "start": 1196.8833798076926, "end": 1197.0233798076924, "probability": 0.99951171875 }, { "word": "复", "start": 1197.0233798076924, "end": 1197.1833798076925, "probability": 0.8349609375 }, { "word": "用", "start": 1197.1833798076925, "end": 1197.3033798076924, "probability": 0.9970703125 }, { "word": "编", "start": 1197.3033798076924, "end": 1197.4233798076925, "probability": 0.9931640625 }, { "word": "码", "start": 1197.4233798076925, "end": 1197.5633798076924, "probability": 1.0 }, { "word": "器", "start": 1197.5633798076924, "end": 1197.7233798076925, "probability": 0.99951171875 }, { "word": "生", "start": 1197.7233798076925, "end": 1197.9033798076925, "probability": 0.990234375 }, { "word": "成", "start": 1197.9033798076925, "end": 1198.0833798076924, "probability": 0.99462890625 }, { "word": "的", "start": 1198.0833798076924, "end": 1198.2033798076925, "probability": 0.99609375 }, { "word": "全", "start": 1198.2033798076925, "end": 1198.5033798076925, "probability": 0.9892578125 }, { "word": "局", "start": 1198.5033798076925, "end": 1198.8233798076924, "probability": 0.99755859375 }, { "word": "K", "start": 1198.8233798076924, "end": 1199.0233798076924, "probability": 0.97265625 }, { "word": "V", "start": 1199.0233798076924, "end": 1199.2233798076925, "probability": 0.99853515625 }, { "word": "了", "start": 1199.2233798076925, "end": 1199.3633798076926, "probability": 0.94189453125 } ], "asr_metrics": { "avg_logprob": -0.08884044880038695, "compression_ratio": 1.2506666666666666, "no_speech_prob": 3.118513780542287e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这里的数字已表示主KV不会进行Token维度的压缩我个人的理解是这大概率是为了保留完整的Token级分辨率毕竟解码器已经在跨层大规模复用编码器生成的全局KV了但即便采用了CSAR附用重索引层仍然需要重新扫描整个百万级Token的上下文因此DeepSeek加入了分层吸收索引器避免这类搜索再次产生过高的算力开销首个完整层会扫描全部内容构建一个最多包含约1.6万" } ], "id": 560, "review_annotation": "", "row_id": 560, "bounded_row_id": 560, "parent_row_ids": [ 560 ], "parent_row_id": 560, "packet_id": 43, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1185.4833798076925, "packet_end": 1214.6713653846155, "raw_text": "毕竟解码器已经在跨层大规模复用编码器生成的全局KV了", "selected_text": "毕竟解码器已经在跨层大规模复用编码器生成的全局KV了", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 560 ], "packet_id": 43, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1199.3633798076926, "end": 1202.2633798076924, "text": "但即便采用了CSAR附用重索引层", "raw_asr_text": "但即便采用了CSAR附用重索引层", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 43, "words": [ { "word": "但", "start": 1199.3633798076926, "end": 1199.6033798076924, "probability": 0.98828125 }, { "word": "即", "start": 1199.6033798076924, "end": 1199.8633798076926, "probability": 0.9990234375 }, { "word": "便", "start": 1199.8633798076926, "end": 1200.0433798076924, "probability": 0.99951171875 }, { "word": "采", "start": 1200.0433798076924, "end": 1200.3033798076924, "probability": 0.9990234375 }, { "word": "用", "start": 1200.3033798076924, "end": 1200.4233798076925, "probability": 0.99951171875 }, { "word": "了", "start": 1200.4233798076925, "end": 1200.5433798076924, "probability": 0.99072265625 }, { "word": "CS", "start": 1200.5433798076924, "end": 1200.7633798076924, "probability": 0.6865234375 }, { "word": "AR", "start": 1200.7633798076924, "end": 1201.0033798076925, "probability": 0.9169921875 }, { "word": "附", "start": 1201.0033798076925, "end": 1201.2833798076924, "probability": 0.8994140625 }, { "word": "用", "start": 1201.2833798076924, "end": 1201.5033798076925, "probability": 0.9970703125 }, { "word": "重", "start": 1201.5033798076925, "end": 1201.7033798076925, "probability": 0.94140625 }, { "word": "索", "start": 1201.7033798076925, "end": 1201.9233798076925, "probability": 0.861328125 }, { "word": "引", "start": 1201.9233798076925, "end": 1202.0833798076924, "probability": 0.98876953125 }, { "word": "层", "start": 1202.0833798076924, "end": 1202.2633798076924, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.08884044880038695, "compression_ratio": 1.2506666666666666, "no_speech_prob": 3.118513780542287e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这里的数字已表示主KV不会进行Token维度的压缩我个人的理解是这大概率是为了保留完整的Token级分辨率毕竟解码器已经在跨层大规模复用编码器生成的全局KV了但即便采用了CSAR附用重索引层仍然需要重新扫描整个百万级Token的上下文因此DeepSeek加入了分层吸收索引器避免这类搜索再次产生过高的算力开销首个完整层会扫描全部内容构建一个最多包含约1.6万" } ], "id": 561, "review_annotation": "", "row_id": 561, "bounded_row_id": 561, "parent_row_ids": [ 561 ], "parent_row_id": 561, "packet_id": 43, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1185.4833798076925, "packet_end": 1214.6713653846155, "raw_text": "但即便采用了CSAR附用重索引层", "selected_text": "但即便采用了CSAR附用重索引层", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 561 ], "packet_id": 43, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1202.2633798076924, "end": 1205.4433798076925, "text": "仍然需要重新扫描整个百万级Token的上下文", "raw_asr_text": "仍然需要重新扫描整个百万级Token的上下文", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 43, "words": [ { "word": "仍", "start": 1202.2633798076924, "end": 1202.5433798076924, "probability": 0.994140625 }, { "word": "然", "start": 1202.5433798076924, "end": 1202.6633798076925, "probability": 0.9990234375 }, { "word": "需要", "start": 1202.6633798076925, "end": 1202.9033798076925, "probability": 0.99267578125 }, { "word": "重", "start": 1202.9033798076925, "end": 1203.1433798076926, "probability": 0.99853515625 }, { "word": "新", "start": 1203.1433798076926, "end": 1203.3033798076924, "probability": 0.99951171875 }, { "word": "扫", "start": 1203.3033798076924, "end": 1203.4833798076925, "probability": 0.9970703125 }, { "word": "描", "start": 1203.4833798076925, "end": 1203.6633798076925, "probability": 0.998046875 }, { "word": "整", "start": 1203.6633798076925, "end": 1203.8833798076926, "probability": 0.99658203125 }, { "word": "个", "start": 1203.8833798076926, "end": 1204.0833798076924, "probability": 0.99951171875 }, { "word": "百", "start": 1204.0833798076924, "end": 1204.2633798076924, "probability": 0.990234375 }, { "word": "万", "start": 1204.2633798076924, "end": 1204.4433798076925, "probability": 0.99951171875 }, { "word": "级", "start": 1204.4433798076925, "end": 1204.6233798076926, "probability": 0.978515625 }, { "word": "T", "start": 1204.6233798076926, "end": 1204.7633798076924, "probability": 0.9638671875 }, { "word": "oken", "start": 1204.7633798076924, "end": 1204.9233798076925, "probability": 0.99853515625 }, { "word": "的", "start": 1204.9233798076925, "end": 1205.0033798076925, "probability": 0.98583984375 }, { "word": "上", "start": 1205.0033798076925, "end": 1205.1233798076926, "probability": 0.99951171875 }, { "word": "下", "start": 1205.1233798076926, "end": 1205.2433798076925, "probability": 0.99951171875 }, { "word": "文", "start": 1205.2433798076925, "end": 1205.4433798076925, "probability": 0.900390625 } ], "asr_metrics": { "avg_logprob": -0.08884044880038695, "compression_ratio": 1.2506666666666666, "no_speech_prob": 3.118513780542287e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这里的数字已表示主KV不会进行Token维度的压缩我个人的理解是这大概率是为了保留完整的Token级分辨率毕竟解码器已经在跨层大规模复用编码器生成的全局KV了但即便采用了CSAR附用重索引层仍然需要重新扫描整个百万级Token的上下文因此DeepSeek加入了分层吸收索引器避免这类搜索再次产生过高的算力开销首个完整层会扫描全部内容构建一个最多包含约1.6万" } ], "id": 562, "review_annotation": "", "row_id": 562, "bounded_row_id": 562, "parent_row_ids": [ 562 ], "parent_row_id": 562, "packet_id": 43, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1185.4833798076925, "packet_end": 1214.6713653846155, "raw_text": "仍然需要重新扫描整个百万级Token的上下文", "selected_text": "仍然需要重新扫描整个百万级Token的上下文", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 562 ], "packet_id": 43, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1205.4433798076925, "end": 1207.9633798076925, "text": "因此DeepSeek加入了分层吸收索引器", "raw_asr_text": "因此DeepSeek加入了分层吸收索引器", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 43, "words": [ { "word": "因", "start": 1205.4433798076925, "end": 1205.7033798076925, "probability": 0.99462890625 }, { "word": "此", "start": 1205.7033798076925, "end": 1205.8633798076926, "probability": 1.0 }, { "word": "De", "start": 1205.8633798076926, "end": 1206.0033798076925, "probability": 0.82470703125 }, { "word": "ep", "start": 1206.0033798076925, "end": 1206.1033798076924, "probability": 0.9228515625 }, { "word": "Se", "start": 1206.1033798076924, "end": 1206.2433798076925, "probability": 0.385986328125 }, { "word": "ek", "start": 1206.2433798076925, "end": 1206.3433798076924, "probability": 0.480712890625 }, { "word": "加入", "start": 1206.3433798076924, "end": 1206.6633798076925, "probability": 0.986328125 }, { "word": "了", "start": 1206.6633798076925, "end": 1206.8233798076924, "probability": 0.9990234375 }, { "word": "分", "start": 1206.8233798076924, "end": 1207.0233798076924, "probability": 0.9892578125 }, { "word": "层", "start": 1207.0233798076924, "end": 1207.2033798076925, "probability": 0.99951171875 }, { "word": "吸", "start": 1207.2033798076925, "end": 1207.3433798076924, "probability": 0.80810546875 }, { "word": "收", "start": 1207.3433798076924, "end": 1207.5033798076925, "probability": 0.51953125 }, { "word": "索", "start": 1207.5033798076925, "end": 1207.6633798076925, "probability": 0.98779296875 }, { "word": "引", "start": 1207.6633798076925, "end": 1207.7833798076924, "probability": 1.0 }, { "word": "器", "start": 1207.7833798076924, "end": 1207.9633798076925, "probability": 0.70849609375 } ], "asr_metrics": { "avg_logprob": -0.08884044880038695, "compression_ratio": 1.2506666666666666, "no_speech_prob": 3.118513780542287e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这里的数字已表示主KV不会进行Token维度的压缩我个人的理解是这大概率是为了保留完整的Token级分辨率毕竟解码器已经在跨层大规模复用编码器生成的全局KV了但即便采用了CSAR附用重索引层仍然需要重新扫描整个百万级Token的上下文因此DeepSeek加入了分层吸收索引器避免这类搜索再次产生过高的算力开销首个完整层会扫描全部内容构建一个最多包含约1.6万" } ], "id": 563, "review_annotation": "", "row_id": 563, "bounded_row_id": 563, "parent_row_ids": [ 563 ], "parent_row_id": 563, "packet_id": 43, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1185.4833798076925, "packet_end": 1214.6713653846155, "raw_text": "因此DeepSeek加入了分层吸收索引器", "selected_text": "因此DeepSeek加入了分层吸收索引器", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 563 ], "packet_id": 43, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1207.9633798076925, "end": 1210.6033798076924, "text": "避免这类搜索再次产生过高的算力开销", "raw_asr_text": "避免这类搜索再次产生过高的算力开销", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 43, "words": [ { "word": "避", "start": 1207.9633798076925, "end": 1208.1833798076925, "probability": 0.9921875 }, { "word": "免", "start": 1208.1833798076925, "end": 1208.3233798076924, "probability": 1.0 }, { "word": "这", "start": 1208.3233798076924, "end": 1208.4433798076925, "probability": 0.9970703125 }, { "word": "类", "start": 1208.4433798076925, "end": 1208.5633798076924, "probability": 0.99951171875 }, { "word": "搜", "start": 1208.5633798076924, "end": 1208.7433798076925, "probability": 0.998046875 }, { "word": "索", "start": 1208.7433798076925, "end": 1208.8633798076926, "probability": 1.0 }, { "word": "再", "start": 1208.8633798076926, "end": 1209.0833798076924, "probability": 0.99853515625 }, { "word": "次", "start": 1209.0833798076924, "end": 1209.2633798076924, "probability": 0.9970703125 }, { "word": "产", "start": 1209.2633798076924, "end": 1209.4033798076925, "probability": 0.9990234375 }, { "word": "生", "start": 1209.4033798076925, "end": 1209.5433798076924, "probability": 1.0 }, { "word": "过", "start": 1209.5433798076924, "end": 1209.7433798076925, "probability": 0.9951171875 }, { "word": "高", "start": 1209.7433798076925, "end": 1209.9233798076925, "probability": 0.9990234375 }, { "word": "的", "start": 1209.9233798076925, "end": 1210.0433798076924, "probability": 0.99755859375 }, { "word": "算", "start": 1210.0433798076924, "end": 1210.1833798076925, "probability": 0.99755859375 }, { "word": "力", "start": 1210.1833798076925, "end": 1210.2833798076924, "probability": 0.9765625 }, { "word": "开", "start": 1210.2833798076924, "end": 1210.4433798076925, "probability": 0.99951171875 }, { "word": "销", "start": 1210.4433798076925, "end": 1210.6033798076924, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.08884044880038695, "compression_ratio": 1.2506666666666666, "no_speech_prob": 3.118513780542287e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这里的数字已表示主KV不会进行Token维度的压缩我个人的理解是这大概率是为了保留完整的Token级分辨率毕竟解码器已经在跨层大规模复用编码器生成的全局KV了但即便采用了CSAR附用重索引层仍然需要重新扫描整个百万级Token的上下文因此DeepSeek加入了分层吸收索引器避免这类搜索再次产生过高的算力开销首个完整层会扫描全部内容构建一个最多包含约1.6万" } ], "id": 564, "review_annotation": "", "row_id": 564, "bounded_row_id": 564, "parent_row_ids": [ 564 ], "parent_row_id": 564, "packet_id": 43, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1185.4833798076925, "packet_end": 1214.6713653846155, "raw_text": "避免这类搜索再次产生过高的算力开销", "selected_text": "避免这类搜索再次产生过高的算力开销", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 564 ], "packet_id": 43, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1210.6033798076924, "end": 1212.4233798076925, "text": "首个完整层会扫描全部内容", "raw_asr_text": "首个完整层会扫描全部内容", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 43, "words": [ { "word": "首", "start": 1210.6033798076924, "end": 1210.8033798076924, "probability": 0.97998046875 }, { "word": "个", "start": 1210.8033798076924, "end": 1210.9433798076925, "probability": 0.99609375 }, { "word": "完", "start": 1210.9433798076925, "end": 1211.0833798076924, "probability": 0.9990234375 }, { "word": "整", "start": 1211.0833798076924, "end": 1211.2033798076925, "probability": 1.0 }, { "word": "层", "start": 1211.2033798076925, "end": 1211.3433798076924, "probability": 0.99951171875 }, { "word": "会", "start": 1211.3433798076924, "end": 1211.4833798076925, "probability": 0.9931640625 }, { "word": "扫", "start": 1211.4833798076925, "end": 1211.6233798076926, "probability": 0.9990234375 }, { "word": "描", "start": 1211.6233798076926, "end": 1211.7633798076924, "probability": 1.0 }, { "word": "全部", "start": 1211.7633798076924, "end": 1212.0633798076924, "probability": 0.99365234375 }, { "word": "内", "start": 1212.0633798076924, "end": 1212.2233798076925, "probability": 0.99365234375 }, { "word": "容", "start": 1212.2233798076925, "end": 1212.4233798076925, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.08884044880038695, "compression_ratio": 1.2506666666666666, "no_speech_prob": 3.118513780542287e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这里的数字已表示主KV不会进行Token维度的压缩我个人的理解是这大概率是为了保留完整的Token级分辨率毕竟解码器已经在跨层大规模复用编码器生成的全局KV了但即便采用了CSAR附用重索引层仍然需要重新扫描整个百万级Token的上下文因此DeepSeek加入了分层吸收索引器避免这类搜索再次产生过高的算力开销首个完整层会扫描全部内容构建一个最多包含约1.6万" } ], "id": 565, "review_annotation": "", "row_id": 565, "bounded_row_id": 565, "parent_row_ids": [ 565 ], "parent_row_id": 565, "packet_id": 43, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1185.4833798076925, "packet_end": 1214.6713653846155, "raw_text": "首个完整层会扫描全部内容", "selected_text": "首个完整层会扫描全部内容", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 565 ], "packet_id": 43, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1212.4233798076925, "end": 1213.8713653846155, "text": "构建一个最多包含约1.6万", "raw_asr_text": "构建一个最多包含约1.6万", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 43, "words": [ { "word": "构", "start": 1212.4233798076925, "end": 1212.7433798076925, "probability": 0.9912109375 }, { "word": "建", "start": 1212.7433798076925, "end": 1212.8833798076926, "probability": 0.99169921875 }, { "word": "一个", "start": 1212.8833798076926, "end": 1213.0433798076924, "probability": 0.99658203125 }, { "word": "最", "start": 1213.0433798076924, "end": 1213.2433798076925, "probability": 0.9990234375 }, { "word": "多", "start": 1213.2433798076925, "end": 1213.4033798076925, "probability": 0.99951171875 }, { "word": "包", "start": 1213.4033798076925, "end": 1213.5433798076924, "probability": 0.9990234375 }, { "word": "含", "start": 1213.5433798076924, "end": 1213.7433798076925, "probability": 0.9873046875 }, { "word": "约", "start": 1213.7433798076925, "end": 1213.9433798076925, "probability": 0.9892578125 }, { "word": "1.", "start": 1213.9433798076925, "end": 1214.2633798076924, "probability": 0.99267578125 }, { "word": "6", "start": 1214.3833798076926, "end": 1214.6433798076926, "probability": 0.9697265625 }, { "word": "万", "start": 1214.6433798076926, "end": 1214.6433798076926, "probability": 0.487060546875 } ], "asr_metrics": { "avg_logprob": -0.08884044880038695, "compression_ratio": 1.2506666666666666, "no_speech_prob": 3.118513780542287e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "这里的数字已表示主KV不会进行Token维度的压缩我个人的理解是这大概率是为了保留完整的Token级分辨率毕竟解码器已经在跨层大规模复用编码器生成的全局KV了但即便采用了CSAR附用重索引层仍然需要重新扫描整个百万级Token的上下文因此DeepSeek加入了分层吸收索引器避免这类搜索再次产生过高的算力开销首个完整层会扫描全部内容构建一个最多包含约1.6万" } ], "id": 566, "review_annotation": "", "row_id": 566, "bounded_row_id": 566, "parent_row_ids": [ 566 ], "parent_row_id": 566, "packet_id": 43, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1185.4833798076925, "packet_end": 1214.6713653846155, "raw_text": "构建一个最多包含约1.6万", "selected_text": "构建一个最多包含约1.6万", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 566 ], "packet_id": 43, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1213.8713653846155, "end": 1215.6913653846154, "text": "最多包含约1.6万个位置的候选池", "raw_asr_text": "最多包含约1.6万个位置的候选池", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 44, "words": [ { "word": "最", "start": 1213.0713653846155, "end": 1213.2113653846156, "probability": 0.5634765625 }, { "word": "多", "start": 1213.2113653846156, "end": 1213.4113653846155, "probability": 0.99658203125 }, { "word": "包", "start": 1213.4113653846155, "end": 1213.5713653846155, "probability": 0.99560546875 }, { "word": "含", "start": 1213.5713653846155, "end": 1213.7513653846156, "probability": 0.998046875 }, { "word": "约", "start": 1213.7513653846156, "end": 1213.9513653846157, "probability": 0.9765625 }, { "word": "1.", "start": 1213.9513653846157, "end": 1214.2513653846156, "probability": 0.98046875 }, { "word": "6", "start": 1214.3713653846155, "end": 1214.5913653846155, "probability": 0.97900390625 }, { "word": "万", "start": 1214.5913653846155, "end": 1214.7513653846156, "probability": 0.9951171875 }, { "word": "个", "start": 1214.7513653846156, "end": 1214.8713653846155, "probability": 0.99365234375 }, { "word": "位", "start": 1214.8713653846155, "end": 1214.9913653846156, "probability": 0.994140625 }, { "word": "置", "start": 1214.9913653846156, "end": 1215.1113653846155, "probability": 0.99853515625 }, { "word": "的", "start": 1215.1113653846155, "end": 1215.2113653846156, "probability": 0.98974609375 }, { "word": "候", "start": 1215.2113653846156, "end": 1215.3313653846155, "probability": 0.93408203125 }, { "word": "选", "start": 1215.3313653846155, "end": 1215.4513653846157, "probability": 1.0 }, { "word": "池", "start": 1215.4513653846157, "end": 1215.6913653846154, "probability": 0.978515625 } ], "asr_metrics": { "avg_logprob": -0.13287843216297238, "compression_ratio": 1.2628726287262872, "no_speech_prob": 2.6276470582731193e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最多包含约1.6万个位置的候选池后续的重锁隐层仅在该候选池内检索筛选出各自排名前512的位置这些层加起来共有20层处理结果会被输入Spark模型架构部分的内容就到这里训练方面也有几处很有意思的改动首先 DeepSeek 没有先训练常规的稠密注意力模型再再之后将其转换为稀疏架构V4.1从0开始直接采用稀疏注意力训练初始上下文长度为6.4万完全没有做稠密" } ], "id": 567, "review_annotation": "", "row_id": 567, "bounded_row_id": 567, "parent_row_ids": [ 567 ], "parent_row_id": 567, "packet_id": 44, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1213.0713653846155, "packet_end": 1242.2593509615385, "raw_text": "最多包含约1.6万个位置的候选池", "selected_text": "最多包含约1.6万个位置的候选池", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 567 ], "packet_id": 44, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1215.6913653846154, "end": 1218.8513653846155, "text": "后续的重锁隐层仅在该候选池内检索", "raw_asr_text": "后续的重锁隐层仅在该候选池内检索", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 44, "words": [ { "word": "后", "start": 1215.6913653846154, "end": 1216.0913653846155, "probability": 0.97412109375 }, { "word": "续", "start": 1216.0913653846155, "end": 1216.2513653846156, "probability": 0.99853515625 }, { "word": "的", "start": 1216.2513653846156, "end": 1216.3713653846155, "probability": 0.984375 }, { "word": "重", "start": 1216.3713653846155, "end": 1216.5313653846156, "probability": 0.99755859375 }, { "word": "锁", "start": 1216.5313653846156, "end": 1216.6913653846154, "probability": 0.8740234375 }, { "word": "隐", "start": 1216.6913653846154, "end": 1216.8313653846155, "probability": 0.71533203125 }, { "word": "层", "start": 1216.8313653846155, "end": 1217.0113653846156, "probability": 0.986328125 }, { "word": "仅", "start": 1217.0113653846156, "end": 1217.2313653846156, "probability": 0.96728515625 }, { "word": "在", "start": 1217.2313653846156, "end": 1217.4513653846157, "probability": 0.998046875 }, { "word": "该", "start": 1217.4513653846157, "end": 1217.6713653846155, "probability": 0.990234375 }, { "word": "候", "start": 1217.6713653846155, "end": 1217.8913653846155, "probability": 0.99853515625 }, { "word": "选", "start": 1217.8913653846155, "end": 1218.0513653846156, "probability": 1.0 }, { "word": "池", "start": 1218.0513653846156, "end": 1218.2513653846156, "probability": 1.0 }, { "word": "内", "start": 1218.2513653846156, "end": 1218.4113653846155, "probability": 0.994140625 }, { "word": "检", "start": 1218.4113653846155, "end": 1218.6313653846155, "probability": 0.94775390625 }, { "word": "索", "start": 1218.6313653846155, "end": 1218.8513653846155, "probability": 0.9130859375 } ], "asr_metrics": { "avg_logprob": -0.13287843216297238, "compression_ratio": 1.2628726287262872, "no_speech_prob": 2.6276470582731193e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最多包含约1.6万个位置的候选池后续的重锁隐层仅在该候选池内检索筛选出各自排名前512的位置这些层加起来共有20层处理结果会被输入Spark模型架构部分的内容就到这里训练方面也有几处很有意思的改动首先 DeepSeek 没有先训练常规的稠密注意力模型再再之后将其转换为稀疏架构V4.1从0开始直接采用稀疏注意力训练初始上下文长度为6.4万完全没有做稠密" } ], "id": 568, "review_annotation": "", "row_id": 568, "bounded_row_id": 568, "parent_row_ids": [ 568 ], "parent_row_id": 568, "packet_id": 44, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1213.0713653846155, "packet_end": 1242.2593509615385, "raw_text": "后续的重锁隐层仅在该候选池内检索", "selected_text": "后续的重锁隐层仅在该候选池内检索", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 568 ], "packet_id": 44, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1218.8513653846155, "end": 1220.9113653846155, "text": "筛选出各自排名前512的位置", "raw_asr_text": "筛选出各自排名前512的位置", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 44, "words": [ { "word": "筛", "start": 1218.8513653846155, "end": 1219.0313653846156, "probability": 0.81005859375 }, { "word": "选", "start": 1219.0313653846156, "end": 1219.1913653846154, "probability": 1.0 }, { "word": "出", "start": 1219.1913653846154, "end": 1219.3313653846155, "probability": 0.994140625 }, { "word": "各", "start": 1219.3313653846155, "end": 1219.4713653846156, "probability": 0.98681640625 }, { "word": "自", "start": 1219.4713653846156, "end": 1219.6113653846155, "probability": 0.9970703125 }, { "word": "排", "start": 1219.6113653846155, "end": 1219.7313653846156, "probability": 0.9990234375 }, { "word": "名", "start": 1219.7313653846156, "end": 1219.8913653846155, "probability": 0.99560546875 }, { "word": "前", "start": 1219.8913653846155, "end": 1220.0713653846155, "probability": 0.9990234375 }, { "word": "5", "start": 1220.0713653846155, "end": 1220.2513653846156, "probability": 0.98046875 }, { "word": "12", "start": 1220.2513653846156, "end": 1220.4713653846156, "probability": 0.96142578125 }, { "word": "的", "start": 1220.4713653846156, "end": 1220.6313653846155, "probability": 0.9169921875 }, { "word": "位", "start": 1220.6313653846155, "end": 1220.7713653846156, "probability": 0.9990234375 }, { "word": "置", "start": 1220.7713653846156, "end": 1220.9113653846155, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.13287843216297238, "compression_ratio": 1.2628726287262872, "no_speech_prob": 2.6276470582731193e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最多包含约1.6万个位置的候选池后续的重锁隐层仅在该候选池内检索筛选出各自排名前512的位置这些层加起来共有20层处理结果会被输入Spark模型架构部分的内容就到这里训练方面也有几处很有意思的改动首先 DeepSeek 没有先训练常规的稠密注意力模型再再之后将其转换为稀疏架构V4.1从0开始直接采用稀疏注意力训练初始上下文长度为6.4万完全没有做稠密" } ], "id": 569, "review_annotation": "", "row_id": 569, "bounded_row_id": 569, "parent_row_ids": [ 569 ], "parent_row_id": 569, "packet_id": 44, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1213.0713653846155, "packet_end": 1242.2593509615385, "raw_text": "筛选出各自排名前512的位置", "selected_text": "筛选出各自排名前512的位置", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 569 ], "packet_id": 44, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1220.9113653846155, "end": 1224.7313653846156, "text": "这些层加起来共有20层处理结果会被输入Spark", "raw_asr_text": "这些层加起来共有20层处理结果会被输入Spark", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 44, "words": [ { "word": "这", "start": 1220.9113653846155, "end": 1221.1313653846155, "probability": 0.99072265625 }, { "word": "些", "start": 1221.1313653846155, "end": 1221.2513653846156, "probability": 0.9990234375 }, { "word": "层", "start": 1221.2513653846156, "end": 1221.4313653846154, "probability": 0.998046875 }, { "word": "加", "start": 1221.4313653846154, "end": 1221.5913653846155, "probability": 0.9951171875 }, { "word": "起", "start": 1221.5913653846155, "end": 1221.7513653846156, "probability": 0.9990234375 }, { "word": "来", "start": 1221.7513653846156, "end": 1221.9313653846154, "probability": 0.99853515625 }, { "word": "共", "start": 1221.9313653846154, "end": 1222.0913653846155, "probability": 0.978515625 }, { "word": "有", "start": 1222.0913653846155, "end": 1222.3113653846156, "probability": 0.998046875 }, { "word": "20", "start": 1222.3113653846156, "end": 1222.6313653846155, "probability": 0.9873046875 }, { "word": "层", "start": 1222.6313653846155, "end": 1222.9113653846155, "probability": 0.9970703125 }, { "word": "处", "start": 1222.9113653846155, "end": 1223.1113653846155, "probability": 0.9892578125 }, { "word": "理", "start": 1223.1113653846155, "end": 1223.2313653846156, "probability": 1.0 }, { "word": "结", "start": 1223.2313653846156, "end": 1223.3913653846155, "probability": 0.99560546875 }, { "word": "果", "start": 1223.3913653846155, "end": 1223.5713653846155, "probability": 1.0 }, { "word": "会", "start": 1223.5713653846155, "end": 1223.7513653846156, "probability": 0.9931640625 }, { "word": "被", "start": 1223.7513653846156, "end": 1223.9113653846155, "probability": 0.88232421875 }, { "word": "输", "start": 1223.9113653846155, "end": 1224.1313653846155, "probability": 0.9990234375 }, { "word": "入", "start": 1224.1313653846155, "end": 1224.3313653846155, "probability": 0.99951171875 }, { "word": "Sp", "start": 1224.3313653846155, "end": 1224.5113653846156, "probability": 0.310302734375 }, { "word": "ark", "start": 1224.5113653846156, "end": 1224.7313653846156, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.13287843216297238, "compression_ratio": 1.2628726287262872, "no_speech_prob": 2.6276470582731193e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最多包含约1.6万个位置的候选池后续的重锁隐层仅在该候选池内检索筛选出各自排名前512的位置这些层加起来共有20层处理结果会被输入Spark模型架构部分的内容就到这里训练方面也有几处很有意思的改动首先 DeepSeek 没有先训练常规的稠密注意力模型再再之后将其转换为稀疏架构V4.1从0开始直接采用稀疏注意力训练初始上下文长度为6.4万完全没有做稠密" } ], "id": 570, "review_annotation": "", "row_id": 570, "bounded_row_id": 570, "parent_row_ids": [ 570 ], "parent_row_id": 570, "packet_id": 44, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1213.0713653846155, "packet_end": 1242.2593509615385, "raw_text": "这些层加起来共有20层处理结果会被输入Spark", "selected_text": "这些层加起来共有20层处理结果会被输入Spark", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 570 ], "packet_id": 44, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1224.7313653846156, "end": 1226.7513653846156, "text": "模型架构部分的内容就到这里", "raw_asr_text": "模型架构部分的内容就到这里", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 44, "words": [ { "word": "模", "start": 1224.7313653846156, "end": 1225.1913653846154, "probability": 0.984375 }, { "word": "型", "start": 1225.1913653846154, "end": 1225.3913653846155, "probability": 0.99951171875 }, { "word": "架", "start": 1225.3913653846155, "end": 1225.5713653846155, "probability": 0.998046875 }, { "word": "构", "start": 1225.5713653846155, "end": 1225.6913653846154, "probability": 0.998046875 }, { "word": "部分", "start": 1225.6913653846154, "end": 1225.9113653846155, "probability": 0.982421875 }, { "word": "的", "start": 1225.9113653846155, "end": 1226.0113653846156, "probability": 0.96435546875 }, { "word": "内", "start": 1226.0113653846156, "end": 1226.1113653846155, "probability": 0.99951171875 }, { "word": "容", "start": 1226.1113653846155, "end": 1226.2513653846156, "probability": 1.0 }, { "word": "就到", "start": 1226.2513653846156, "end": 1226.4913653846156, "probability": 0.98193359375 }, { "word": "这里", "start": 1226.4913653846156, "end": 1226.7513653846156, "probability": 0.9931640625 } ], "asr_metrics": { "avg_logprob": -0.13287843216297238, "compression_ratio": 1.2628726287262872, "no_speech_prob": 2.6276470582731193e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最多包含约1.6万个位置的候选池后续的重锁隐层仅在该候选池内检索筛选出各自排名前512的位置这些层加起来共有20层处理结果会被输入Spark模型架构部分的内容就到这里训练方面也有几处很有意思的改动首先 DeepSeek 没有先训练常规的稠密注意力模型再再之后将其转换为稀疏架构V4.1从0开始直接采用稀疏注意力训练初始上下文长度为6.4万完全没有做稠密" } ], "id": 571, "review_annotation": "", "row_id": 571, "bounded_row_id": 571, "parent_row_ids": [ 571 ], "parent_row_id": 571, "packet_id": 44, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1213.0713653846155, "packet_end": 1242.2593509615385, "raw_text": "模型架构部分的内容就到这里", "selected_text": "模型架构部分的内容就到这里", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 571 ], "packet_id": 44, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1226.7513653846156, "end": 1229.3313653846155, "text": "训练方面也有几处很有意思的改动", "raw_asr_text": "训练方面也有几处很有意思的改动", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 44, "words": [ { "word": "训", "start": 1226.7513653846156, "end": 1227.1313653846155, "probability": 0.98095703125 }, { "word": "练", "start": 1227.1313653846155, "end": 1227.2713653846156, "probability": 0.99951171875 }, { "word": "方面", "start": 1227.2713653846156, "end": 1227.5313653846156, "probability": 0.99267578125 }, { "word": "也", "start": 1227.5313653846156, "end": 1227.7713653846156, "probability": 0.990234375 }, { "word": "有", "start": 1227.7713653846156, "end": 1227.8913653846155, "probability": 0.99853515625 }, { "word": "几", "start": 1227.8913653846155, "end": 1228.0713653846155, "probability": 0.99951171875 }, { "word": "处", "start": 1228.0713653846155, "end": 1228.2713653846156, "probability": 0.99658203125 }, { "word": "很有", "start": 1228.2713653846156, "end": 1228.5913653846155, "probability": 0.98779296875 }, { "word": "意思", "start": 1228.5913653846155, "end": 1228.8713653846155, "probability": 0.99951171875 }, { "word": "的", "start": 1228.8713653846155, "end": 1228.9713653846156, "probability": 0.98876953125 }, { "word": "改", "start": 1228.9713653846156, "end": 1229.1113653846155, "probability": 0.99462890625 }, { "word": "动", "start": 1229.1113653846155, "end": 1229.3313653846155, "probability": 0.9970703125 } ], "asr_metrics": { "avg_logprob": -0.13287843216297238, "compression_ratio": 1.2628726287262872, "no_speech_prob": 2.6276470582731193e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最多包含约1.6万个位置的候选池后续的重锁隐层仅在该候选池内检索筛选出各自排名前512的位置这些层加起来共有20层处理结果会被输入Spark模型架构部分的内容就到这里训练方面也有几处很有意思的改动首先 DeepSeek 没有先训练常规的稠密注意力模型再再之后将其转换为稀疏架构V4.1从0开始直接采用稀疏注意力训练初始上下文长度为6.4万完全没有做稠密" } ], "id": 572, "review_annotation": "", "row_id": 572, "bounded_row_id": 572, "parent_row_ids": [ 572 ], "parent_row_id": 572, "packet_id": 44, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1213.0713653846155, "packet_end": 1242.2593509615385, "raw_text": "训练方面也有几处很有意思的改动", "selected_text": "训练方面也有几处很有意思的改动", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 572 ], "packet_id": 44, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1229.3313653846155, "end": 1232.8513653846155, "text": "首先 DeepSeek 没有先训练常规的稠密注意力模型", "raw_asr_text": "首先 DeepSeek 没有先训练常规的稠密注意力模型", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 44, "words": [ { "word": "首先", "start": 1229.3313653846155, "end": 1230.0313653846156, "probability": 0.99609375 }, { "word": " Deep", "start": 1230.0313653846156, "end": 1230.3713653846155, "probability": 0.258544921875 }, { "word": "Se", "start": 1230.3713653846155, "end": 1230.5313653846156, "probability": 0.3916015625 }, { "word": "ek", "start": 1230.5313653846156, "end": 1230.6313653846155, "probability": 0.5439453125 }, { "word": " 没有", "start": 1230.8513653846155, "end": 1230.9313653846154, "probability": 0.9921875 }, { "word": "先", "start": 1230.9313653846154, "end": 1231.1513653846155, "probability": 0.99169921875 }, { "word": "训", "start": 1231.1513653846155, "end": 1231.3113653846156, "probability": 0.99609375 }, { "word": "练", "start": 1231.3113653846156, "end": 1231.4313653846154, "probability": 0.9990234375 }, { "word": "常", "start": 1231.4313653846154, "end": 1231.6113653846155, "probability": 0.9814453125 }, { "word": "规", "start": 1231.6113653846155, "end": 1231.8513653846155, "probability": 0.9990234375 }, { "word": "的", "start": 1231.8513653846155, "end": 1231.9513653846157, "probability": 0.99560546875 }, { "word": "稠", "start": 1231.9513653846157, "end": 1232.1113653846155, "probability": 0.970703125 }, { "word": "密", "start": 1232.1113653846155, "end": 1232.2313653846156, "probability": 0.91259765625 }, { "word": "注意", "start": 1232.2313653846156, "end": 1232.3713653846155, "probability": 0.97998046875 }, { "word": "力", "start": 1232.3713653846155, "end": 1232.4913653846156, "probability": 0.99951171875 }, { "word": "模", "start": 1232.4913653846156, "end": 1232.6313653846155, "probability": 0.998046875 }, { "word": "型", "start": 1232.6313653846155, "end": 1232.8513653846155, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.13287843216297238, "compression_ratio": 1.2628726287262872, "no_speech_prob": 2.6276470582731193e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最多包含约1.6万个位置的候选池后续的重锁隐层仅在该候选池内检索筛选出各自排名前512的位置这些层加起来共有20层处理结果会被输入Spark模型架构部分的内容就到这里训练方面也有几处很有意思的改动首先 DeepSeek 没有先训练常规的稠密注意力模型再再之后将其转换为稀疏架构V4.1从0开始直接采用稀疏注意力训练初始上下文长度为6.4万完全没有做稠密" } ], "id": 573, "review_annotation": "", "row_id": 573, "bounded_row_id": 573, "parent_row_ids": [ 573 ], "parent_row_id": 573, "packet_id": 44, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1213.0713653846155, "packet_end": 1242.2593509615385, "raw_text": "首先 DeepSeek 没有先训练常规的稠密注意力模型", "selected_text": "首先 DeepSeek 没有先训练常规的稠密注意力模型", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 573 ], "packet_id": 44, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1232.8513653846155, "end": 1235.2913653846156, "text": "再再之后将其转换为稀疏架构", "raw_asr_text": "再再之后将其转换为稀疏架构", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 44, "words": [ { "word": "再", "start": 1232.8513653846155, "end": 1233.3113653846156, "probability": 0.9345703125 }, { "word": "再", "start": 1233.3113653846156, "end": 1233.4713653846156, "probability": 0.5087890625 }, { "word": "之", "start": 1233.4713653846156, "end": 1233.6913653846154, "probability": 0.990234375 }, { "word": "后", "start": 1233.6913653846154, "end": 1233.8313653846155, "probability": 1.0 }, { "word": "将", "start": 1233.8313653846155, "end": 1234.0513653846156, "probability": 0.95703125 }, { "word": "其", "start": 1234.0513653846156, "end": 1234.2113653846156, "probability": 0.99658203125 }, { "word": "转", "start": 1234.2113653846156, "end": 1234.4113653846155, "probability": 0.998046875 }, { "word": "换", "start": 1234.4113653846155, "end": 1234.5713653846155, "probability": 0.9951171875 }, { "word": "为", "start": 1234.5713653846155, "end": 1234.7113653846156, "probability": 0.986328125 }, { "word": "稀", "start": 1234.7113653846156, "end": 1234.8713653846155, "probability": 0.8466796875 }, { "word": "疏", "start": 1234.8713653846155, "end": 1234.9513653846157, "probability": 0.990234375 }, { "word": "架", "start": 1234.9513653846157, "end": 1235.1313653846155, "probability": 0.9951171875 }, { "word": "构", "start": 1235.1313653846155, "end": 1235.2913653846156, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.13287843216297238, "compression_ratio": 1.2628726287262872, "no_speech_prob": 2.6276470582731193e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最多包含约1.6万个位置的候选池后续的重锁隐层仅在该候选池内检索筛选出各自排名前512的位置这些层加起来共有20层处理结果会被输入Spark模型架构部分的内容就到这里训练方面也有几处很有意思的改动首先 DeepSeek 没有先训练常规的稠密注意力模型再再之后将其转换为稀疏架构V4.1从0开始直接采用稀疏注意力训练初始上下文长度为6.4万完全没有做稠密" } ], "id": 574, "review_annotation": "", "row_id": 574, "bounded_row_id": 574, "parent_row_ids": [ 574 ], "parent_row_id": 574, "packet_id": 44, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1213.0713653846155, "packet_end": 1242.2593509615385, "raw_text": "再再之后将其转换为稀疏架构", "selected_text": "再再之后将其转换为稀疏架构", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 574 ], "packet_id": 44, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1235.2913653846156, "end": 1238.6313653846155, "text": "V4.1从0开始直接采用稀疏注意力训练", "raw_asr_text": "V4.1从0开始直接采用稀疏注意力训练", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 44, "words": [ { "word": "V", "start": 1235.2913653846156, "end": 1235.6313653846155, "probability": 0.85498046875 }, { "word": "4.", "start": 1235.6313653846155, "end": 1235.8513653846155, "probability": 0.95751953125 }, { "word": "1", "start": 1235.9513653846157, "end": 1236.1113653846155, "probability": 0.99560546875 }, { "word": "从", "start": 1236.1113653846155, "end": 1236.3713653846155, "probability": 0.83349609375 }, { "word": "0", "start": 1236.3713653846155, "end": 1236.5313653846156, "probability": 0.74609375 }, { "word": "开始", "start": 1236.5313653846156, "end": 1236.8113653846156, "probability": 0.9892578125 }, { "word": "直接", "start": 1236.8113653846156, "end": 1237.0913653846155, "probability": 0.99169921875 }, { "word": "采", "start": 1237.0913653846155, "end": 1237.2913653846156, "probability": 0.99853515625 }, { "word": "用", "start": 1237.2913653846156, "end": 1237.4513653846157, "probability": 0.99951171875 }, { "word": "稀", "start": 1237.4513653846157, "end": 1237.7113653846156, "probability": 0.974609375 }, { "word": "疏", "start": 1237.7113653846156, "end": 1237.8913653846155, "probability": 1.0 }, { "word": "注意", "start": 1237.8913653846155, "end": 1238.1313653846155, "probability": 0.9931640625 }, { "word": "力", "start": 1238.1313653846155, "end": 1238.2713653846156, "probability": 0.99951171875 }, { "word": "训", "start": 1238.2713653846156, "end": 1238.4513653846157, "probability": 0.99609375 }, { "word": "练", "start": 1238.4513653846157, "end": 1238.6313653846155, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.13287843216297238, "compression_ratio": 1.2628726287262872, "no_speech_prob": 2.6276470582731193e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最多包含约1.6万个位置的候选池后续的重锁隐层仅在该候选池内检索筛选出各自排名前512的位置这些层加起来共有20层处理结果会被输入Spark模型架构部分的内容就到这里训练方面也有几处很有意思的改动首先 DeepSeek 没有先训练常规的稠密注意力模型再再之后将其转换为稀疏架构V4.1从0开始直接采用稀疏注意力训练初始上下文长度为6.4万完全没有做稠密" } ], "id": 575, "review_annotation": "", "row_id": 575, "bounded_row_id": 575, "parent_row_ids": [ 575 ], "parent_row_id": 575, "packet_id": 44, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1213.0713653846155, "packet_end": 1242.2593509615385, "raw_text": "V4.1从0开始直接采用稀疏注意力训练", "selected_text": "V4.1从0开始直接采用稀疏注意力训练", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 575 ], "packet_id": 44, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1238.6313653846155, "end": 1240.8113653846156, "text": "初始上下文长度为6.4万", "raw_asr_text": "初始上下文长度为6.4万", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 44, "words": [ { "word": "初", "start": 1238.6313653846155, "end": 1239.0513653846156, "probability": 0.76708984375 }, { "word": "始", "start": 1239.0513653846156, "end": 1239.1913653846154, "probability": 0.97705078125 }, { "word": "上", "start": 1239.1913653846154, "end": 1239.3513653846155, "probability": 0.994140625 }, { "word": "下", "start": 1239.3513653846155, "end": 1239.4913653846156, "probability": 0.9970703125 }, { "word": "文", "start": 1239.4913653846156, "end": 1239.6313653846155, "probability": 0.64306640625 }, { "word": "长", "start": 1239.6313653846155, "end": 1239.7713653846156, "probability": 0.98974609375 }, { "word": "度", "start": 1239.7713653846156, "end": 1239.9113653846155, "probability": 0.99951171875 }, { "word": "为", "start": 1239.9113653846155, "end": 1240.0713653846155, "probability": 0.99609375 }, { "word": "6.", "start": 1240.0713653846155, "end": 1240.2913653846156, "probability": 0.9541015625 }, { "word": "4", "start": 1240.4113653846155, "end": 1240.6513653846155, "probability": 0.88916015625 }, { "word": "万", "start": 1240.6513653846155, "end": 1240.8113653846156, "probability": 0.99267578125 } ], "asr_metrics": { "avg_logprob": -0.13287843216297238, "compression_ratio": 1.2628726287262872, "no_speech_prob": 2.6276470582731193e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "最多包含约1.6万个位置的候选池后续的重锁隐层仅在该候选池内检索筛选出各自排名前512的位置这些层加起来共有20层处理结果会被输入Spark模型架构部分的内容就到这里训练方面也有几处很有意思的改动首先 DeepSeek 没有先训练常规的稠密注意力模型再再之后将其转换为稀疏架构V4.1从0开始直接采用稀疏注意力训练初始上下文长度为6.4万完全没有做稠密" } ], "id": 576, "review_annotation": "", "row_id": 576, "bounded_row_id": 576, "parent_row_ids": [ 576 ], "parent_row_id": 576, "packet_id": 44, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1213.0713653846155, "packet_end": 1242.2593509615385, "raw_text": "初始上下文长度为6.4万", "selected_text": "初始上下文长度为6.4万", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 576 ], "packet_id": 44, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1241.4593509615386, "end": 1242.8593509615387, "text": "完全没有做稠密注意力预热", "raw_asr_text": "完全没有做稠密注意力预热", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 45, "words": [ { "word": "完全", "start": 1240.6593509615386, "end": 1241.3593509615387, "probability": 0.94482421875 }, { "word": "没有", "start": 1241.3593509615387, "end": 1241.6393509615386, "probability": 0.95361328125 }, { "word": "做", "start": 1241.6393509615386, "end": 1241.8393509615387, "probability": 0.95654296875 }, { "word": "稠", "start": 1241.8393509615387, "end": 1242.0593509615387, "probability": 0.92431640625 }, { "word": "密", "start": 1242.0593509615387, "end": 1242.1993509615386, "probability": 0.9638671875 }, { "word": "注意", "start": 1242.1993509615386, "end": 1242.3593509615387, "probability": 0.98876953125 }, { "word": "力", "start": 1242.3593509615387, "end": 1242.4993509615385, "probability": 0.99853515625 }, { "word": "预", "start": 1242.4993509615385, "end": 1242.6593509615386, "probability": 0.9970703125 }, { "word": "热", "start": 1242.6593509615386, "end": 1242.8593509615387, "probability": 0.9951171875 } ], "asr_metrics": { "avg_logprob": -0.09403635540279733, "compression_ratio": 1.2727272727272727, "no_speech_prob": 5.203060204905796e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "完全没有做稠密注意力预热等训练量达到约34万亿tokens时他们会直接将上下文长度扩展到100万最终在总计45万亿tokens的多莫态数据上完成预训练因此CSAR CED和稀疏检索行为都不是后续才新增的模型实际上是从头学习这些机制的而如果一家实验室能够不依赖稠密注意力从零成功训练出稀疏注意力这通常就意味着这套注意力机制稳定可靠且具备很高的潜力后续" } ], "id": 577, "review_annotation": "", "row_id": 577, "bounded_row_id": 577, "parent_row_ids": [ 577 ], "parent_row_id": 577, "packet_id": 45, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1240.6593509615386, "packet_end": 1269.8473365384616, "raw_text": "完全没有做稠密注意力预热", "selected_text": "完全没有做稠密注意力预热", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 577 ], "packet_id": 45, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1242.8593509615387, "end": 1245.5393509615387, "text": "等训练量达到约34万亿tokens时", "raw_asr_text": "等训练量达到约34万亿tokens时", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 45, "words": [ { "word": "等", "start": 1242.8593509615387, "end": 1243.2193509615386, "probability": 0.89501953125 }, { "word": "训", "start": 1243.2193509615386, "end": 1243.3993509615386, "probability": 0.99658203125 }, { "word": "练", "start": 1243.3993509615386, "end": 1243.5193509615385, "probability": 0.9990234375 }, { "word": "量", "start": 1243.5193509615385, "end": 1243.6593509615386, "probability": 0.99560546875 }, { "word": "达", "start": 1243.6593509615386, "end": 1243.8393509615387, "probability": 0.99853515625 }, { "word": "到", "start": 1243.8393509615387, "end": 1243.9793509615386, "probability": 0.98486328125 }, { "word": "约", "start": 1243.9793509615386, "end": 1244.1393509615386, "probability": 0.9013671875 }, { "word": "34", "start": 1244.1393509615386, "end": 1244.4393509615386, "probability": 0.97998046875 }, { "word": "万", "start": 1244.4393509615386, "end": 1244.7793509615385, "probability": 0.97021484375 }, { "word": "亿", "start": 1244.7793509615385, "end": 1244.8993509615386, "probability": 0.9853515625 }, { "word": "t", "start": 1244.8993509615386, "end": 1245.0393509615387, "probability": 0.486083984375 }, { "word": "ok", "start": 1245.0393509615387, "end": 1245.1593509615386, "probability": 0.470947265625 }, { "word": "ens", "start": 1245.1593509615386, "end": 1245.2993509615387, "probability": 0.98828125 }, { "word": "时", "start": 1245.2993509615387, "end": 1245.5393509615387, "probability": 0.943359375 } ], "asr_metrics": { "avg_logprob": -0.09403635540279733, "compression_ratio": 1.2727272727272727, "no_speech_prob": 5.203060204905796e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "完全没有做稠密注意力预热等训练量达到约34万亿tokens时他们会直接将上下文长度扩展到100万最终在总计45万亿tokens的多莫态数据上完成预训练因此CSAR CED和稀疏检索行为都不是后续才新增的模型实际上是从头学习这些机制的而如果一家实验室能够不依赖稠密注意力从零成功训练出稀疏注意力这通常就意味着这套注意力机制稳定可靠且具备很高的潜力后续" } ], "id": 578, "review_annotation": "", "row_id": 578, "bounded_row_id": 578, "parent_row_ids": [ 578 ], "parent_row_id": 578, "packet_id": 45, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1240.6593509615386, "packet_end": 1269.8473365384616, "raw_text": "等训练量达到约34万亿tokens时", "selected_text": "等训练量达到约34万亿tokens时", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 578 ], "packet_id": 45, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1245.5393509615387, "end": 1248.1193509615387, "text": "他们会直接将上下文长度扩展到100万", "raw_asr_text": "他们会直接将上下文长度扩展到100万", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 45, "words": [ { "word": "他们", "start": 1245.5393509615387, "end": 1245.7993509615387, "probability": 0.73193359375 }, { "word": "会", "start": 1245.7993509615387, "end": 1245.9793509615386, "probability": 0.9951171875 }, { "word": "直接", "start": 1245.9793509615386, "end": 1246.2593509615385, "probability": 0.990234375 }, { "word": "将", "start": 1246.2593509615385, "end": 1246.4393509615386, "probability": 0.9970703125 }, { "word": "上", "start": 1246.4393509615386, "end": 1246.6593509615386, "probability": 0.9912109375 }, { "word": "下", "start": 1246.6593509615386, "end": 1246.7993509615387, "probability": 0.99609375 }, { "word": "文", "start": 1246.7993509615387, "end": 1246.9593509615386, "probability": 0.92529296875 }, { "word": "长", "start": 1246.9593509615386, "end": 1247.0993509615387, "probability": 0.9873046875 }, { "word": "度", "start": 1247.0993509615387, "end": 1247.2593509615385, "probability": 0.99951171875 }, { "word": "扩", "start": 1247.2593509615385, "end": 1247.4793509615386, "probability": 0.9892578125 }, { "word": "展", "start": 1247.4793509615386, "end": 1247.6193509615387, "probability": 0.99658203125 }, { "word": "到", "start": 1247.6193509615387, "end": 1247.7193509615386, "probability": 0.99560546875 }, { "word": "100", "start": 1247.7193509615386, "end": 1247.9393509615386, "probability": 0.943359375 }, { "word": "万", "start": 1247.9393509615386, "end": 1248.1193509615387, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.09403635540279733, "compression_ratio": 1.2727272727272727, "no_speech_prob": 5.203060204905796e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "完全没有做稠密注意力预热等训练量达到约34万亿tokens时他们会直接将上下文长度扩展到100万最终在总计45万亿tokens的多莫态数据上完成预训练因此CSAR CED和稀疏检索行为都不是后续才新增的模型实际上是从头学习这些机制的而如果一家实验室能够不依赖稠密注意力从零成功训练出稀疏注意力这通常就意味着这套注意力机制稳定可靠且具备很高的潜力后续" } ], "id": 579, "review_annotation": "", "row_id": 579, "bounded_row_id": 579, "parent_row_ids": [ 579 ], "parent_row_id": 579, "packet_id": 45, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1240.6593509615386, "packet_end": 1269.8473365384616, "raw_text": "他们会直接将上下文长度扩展到100万", "selected_text": "他们会直接将上下文长度扩展到100万", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 579 ], "packet_id": 45, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1248.1193509615387, "end": 1252.3593509615387, "text": "最终在总计45万亿tokens的多莫态数据上完成预训练", "raw_asr_text": "最终在总计45万亿tokens的多莫态数据上完成预训练", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 45, "words": [ { "word": "最", "start": 1248.1193509615387, "end": 1248.3993509615386, "probability": 0.982421875 }, { "word": "终", "start": 1248.3993509615386, "end": 1248.5593509615387, "probability": 0.99951171875 }, { "word": "在", "start": 1248.5593509615387, "end": 1248.7593509615385, "probability": 0.98486328125 }, { "word": "总", "start": 1248.7593509615385, "end": 1249.0393509615387, "probability": 0.99658203125 }, { "word": "计", "start": 1249.0393509615387, "end": 1249.2393509615385, "probability": 0.98828125 }, { "word": "45", "start": 1249.2393509615385, "end": 1249.9593509615386, "probability": 0.96142578125 }, { "word": "万", "start": 1249.9593509615386, "end": 1250.1393509615386, "probability": 0.99609375 }, { "word": "亿", "start": 1250.1393509615386, "end": 1250.3193509615387, "probability": 0.9990234375 }, { "word": "t", "start": 1250.3193509615387, "end": 1250.4793509615386, "probability": 0.98974609375 }, { "word": "ok", "start": 1250.4793509615386, "end": 1250.5993509615387, "probability": 0.994140625 }, { "word": "ens", "start": 1250.5993509615387, "end": 1250.7593509615385, "probability": 1.0 }, { "word": "的", "start": 1250.7593509615385, "end": 1250.8993509615386, "probability": 0.94189453125 }, { "word": "多", "start": 1250.8993509615386, "end": 1251.0793509615387, "probability": 0.9892578125 }, { "word": "莫", "start": 1251.0793509615387, "end": 1251.1993509615386, "probability": 0.6796875 }, { "word": "态", "start": 1251.1993509615386, "end": 1251.3193509615387, "probability": 0.962890625 }, { "word": "数", "start": 1251.3193509615387, "end": 1251.4593509615386, "probability": 0.994140625 }, { "word": "据", "start": 1251.4593509615386, "end": 1251.5393509615387, "probability": 0.9990234375 }, { "word": "上", "start": 1251.5393509615387, "end": 1251.6793509615386, "probability": 0.998046875 }, { "word": "完成", "start": 1251.6793509615386, "end": 1251.9793509615386, "probability": 0.99658203125 }, { "word": "预", "start": 1251.9793509615386, "end": 1252.0793509615387, "probability": 0.9921875 }, { "word": "训", "start": 1252.0793509615387, "end": 1252.1993509615386, "probability": 1.0 }, { "word": "练", "start": 1252.1993509615386, "end": 1252.3593509615387, "probability": 0.9951171875 } ], "asr_metrics": { "avg_logprob": -0.09403635540279733, "compression_ratio": 1.2727272727272727, "no_speech_prob": 5.203060204905796e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "完全没有做稠密注意力预热等训练量达到约34万亿tokens时他们会直接将上下文长度扩展到100万最终在总计45万亿tokens的多莫态数据上完成预训练因此CSAR CED和稀疏检索行为都不是后续才新增的模型实际上是从头学习这些机制的而如果一家实验室能够不依赖稠密注意力从零成功训练出稀疏注意力这通常就意味着这套注意力机制稳定可靠且具备很高的潜力后续" } ], "id": 580, "review_annotation": "", "row_id": 580, "bounded_row_id": 580, "parent_row_ids": [ 580 ], "parent_row_id": 580, "packet_id": 45, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1240.6593509615386, "packet_end": 1269.8473365384616, "raw_text": "最终在总计45万亿tokens的多莫态数据上完成预训练", "selected_text": "最终在总计45万亿tokens的多莫态数据上完成预训练", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 580 ], "packet_id": 45, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1252.3593509615387, "end": 1256.9793509615386, "text": "因此CSAR CED和稀疏检索行为都不是后续才新增的", "raw_asr_text": "因此CSAR CED和稀疏检索行为都不是后续才新增的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 45, "words": [ { "word": "因", "start": 1252.3593509615387, "end": 1252.6393509615386, "probability": 0.9931640625 }, { "word": "此", "start": 1252.6393509615386, "end": 1252.9193509615386, "probability": 0.99951171875 }, { "word": "CS", "start": 1252.9193509615386, "end": 1253.3193509615387, "probability": 0.444580078125 }, { "word": "AR", "start": 1253.3193509615387, "end": 1253.6993509615386, "probability": 0.95703125 }, { "word": " C", "start": 1253.6993509615386, "end": 1253.9993509615385, "probability": 0.53173828125 }, { "word": "ED", "start": 1253.9993509615385, "end": 1254.2393509615385, "probability": 0.982421875 }, { "word": "和", "start": 1254.2393509615385, "end": 1254.5593509615387, "probability": 0.8798828125 }, { "word": "稀", "start": 1254.5593509615387, "end": 1254.7593509615385, "probability": 0.84375 }, { "word": "疏", "start": 1254.7593509615385, "end": 1254.8993509615386, "probability": 0.9873046875 }, { "word": "检", "start": 1254.8993509615386, "end": 1255.0793509615387, "probability": 0.98583984375 }, { "word": "索", "start": 1255.0793509615387, "end": 1255.2593509615385, "probability": 0.9931640625 }, { "word": "行", "start": 1255.2593509615385, "end": 1255.3993509615386, "probability": 0.99462890625 }, { "word": "为", "start": 1255.3993509615386, "end": 1255.5393509615387, "probability": 0.99560546875 }, { "word": "都", "start": 1255.5393509615387, "end": 1255.6993509615386, "probability": 0.97705078125 }, { "word": "不是", "start": 1255.6993509615386, "end": 1255.9393509615386, "probability": 0.99853515625 }, { "word": "后", "start": 1255.9393509615386, "end": 1256.1793509615386, "probability": 0.99658203125 }, { "word": "续", "start": 1256.1793509615386, "end": 1256.3593509615387, "probability": 0.9990234375 }, { "word": "才", "start": 1256.3593509615387, "end": 1256.4793509615386, "probability": 0.9873046875 }, { "word": "新", "start": 1256.4793509615386, "end": 1256.6593509615386, "probability": 0.99169921875 }, { "word": "增", "start": 1256.6593509615386, "end": 1256.8393509615387, "probability": 1.0 }, { "word": "的", "start": 1256.8393509615387, "end": 1256.9793509615386, "probability": 0.96240234375 } ], "asr_metrics": { "avg_logprob": -0.09403635540279733, "compression_ratio": 1.2727272727272727, "no_speech_prob": 5.203060204905796e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "完全没有做稠密注意力预热等训练量达到约34万亿tokens时他们会直接将上下文长度扩展到100万最终在总计45万亿tokens的多莫态数据上完成预训练因此CSAR CED和稀疏检索行为都不是后续才新增的模型实际上是从头学习这些机制的而如果一家实验室能够不依赖稠密注意力从零成功训练出稀疏注意力这通常就意味着这套注意力机制稳定可靠且具备很高的潜力后续" } ], "id": 581, "review_annotation": "", "row_id": 581, "bounded_row_id": 581, "parent_row_ids": [ 581 ], "parent_row_id": 581, "packet_id": 45, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1240.6593509615386, "packet_end": 1269.8473365384616, "raw_text": "因此CSAR CED和稀疏检索行为都不是后续才新增的", "selected_text": "因此CSAR CED和稀疏检索行为都不是后续才新增的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 581 ], "packet_id": 45, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1256.9793509615386, "end": 1259.5993509615387, "text": "模型实际上是从头学习这些机制的", "raw_asr_text": "模型实际上是从头学习这些机制的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 45, "words": [ { "word": "模", "start": 1256.9793509615386, "end": 1257.1993509615386, "probability": 0.9912109375 }, { "word": "型", "start": 1257.1993509615386, "end": 1257.4193509615386, "probability": 0.99951171875 }, { "word": "实", "start": 1257.4193509615386, "end": 1257.6593509615386, "probability": 0.9912109375 }, { "word": "际", "start": 1257.6593509615386, "end": 1257.8393509615387, "probability": 1.0 }, { "word": "上", "start": 1257.8393509615387, "end": 1258.0193509615385, "probability": 0.9970703125 }, { "word": "是", "start": 1258.0193509615385, "end": 1258.1993509615386, "probability": 0.99169921875 }, { "word": "从", "start": 1258.1993509615386, "end": 1258.3793509615386, "probability": 0.9970703125 }, { "word": "头", "start": 1258.3793509615386, "end": 1258.5393509615387, "probability": 0.99462890625 }, { "word": "学", "start": 1258.5393509615387, "end": 1258.7193509615386, "probability": 0.99658203125 }, { "word": "习", "start": 1258.7193509615386, "end": 1258.8793509615386, "probability": 0.99951171875 }, { "word": "这", "start": 1258.8793509615386, "end": 1258.9993509615385, "probability": 0.99560546875 }, { "word": "些", "start": 1258.9993509615385, "end": 1259.1193509615387, "probability": 0.99951171875 }, { "word": "机", "start": 1259.1193509615387, "end": 1259.2993509615387, "probability": 0.998046875 }, { "word": "制", "start": 1259.2993509615387, "end": 1259.4393509615386, "probability": 0.97216796875 }, { "word": "的", "start": 1259.4393509615386, "end": 1259.5993509615387, "probability": 0.994140625 } ], "asr_metrics": { "avg_logprob": -0.09403635540279733, "compression_ratio": 1.2727272727272727, "no_speech_prob": 5.203060204905796e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "完全没有做稠密注意力预热等训练量达到约34万亿tokens时他们会直接将上下文长度扩展到100万最终在总计45万亿tokens的多莫态数据上完成预训练因此CSAR CED和稀疏检索行为都不是后续才新增的模型实际上是从头学习这些机制的而如果一家实验室能够不依赖稠密注意力从零成功训练出稀疏注意力这通常就意味着这套注意力机制稳定可靠且具备很高的潜力后续" } ], "id": 582, "review_annotation": "", "row_id": 582, "bounded_row_id": 582, "parent_row_ids": [ 582 ], "parent_row_id": 582, "packet_id": 45, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1240.6593509615386, "packet_end": 1269.8473365384616, "raw_text": "模型实际上是从头学习这些机制的", "selected_text": "模型实际上是从头学习这些机制的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 582 ], "packet_id": 45, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1259.5993509615387, "end": 1262.1993509615386, "text": "而如果一家实验室能够不依赖稠密注意力", "raw_asr_text": "而如果一家实验室能够不依赖稠密注意力", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 45, "words": [ { "word": "而", "start": 1259.5993509615387, "end": 1259.8593509615387, "probability": 0.9501953125 }, { "word": "如果", "start": 1259.8593509615387, "end": 1260.0593509615387, "probability": 0.9931640625 }, { "word": "一", "start": 1260.0593509615387, "end": 1260.1993509615386, "probability": 0.99658203125 }, { "word": "家", "start": 1260.1993509615386, "end": 1260.3193509615387, "probability": 0.9873046875 }, { "word": "实", "start": 1260.3193509615387, "end": 1260.5193509615385, "probability": 0.99658203125 }, { "word": "验", "start": 1260.5193509615385, "end": 1260.6393509615386, "probability": 0.99951171875 }, { "word": "室", "start": 1260.6393509615386, "end": 1260.8193509615387, "probability": 0.998046875 }, { "word": "能", "start": 1260.8193509615387, "end": 1260.9793509615386, "probability": 0.9892578125 }, { "word": "够", "start": 1260.9793509615386, "end": 1261.0593509615387, "probability": 0.9892578125 }, { "word": "不", "start": 1261.0593509615387, "end": 1261.2793509615385, "probability": 0.9970703125 }, { "word": "依", "start": 1261.2793509615385, "end": 1261.4393509615386, "probability": 0.9990234375 }, { "word": "赖", "start": 1261.4393509615386, "end": 1261.5593509615387, "probability": 0.9970703125 }, { "word": "稠", "start": 1261.5593509615387, "end": 1261.6993509615386, "probability": 0.99853515625 }, { "word": "密", "start": 1261.6993509615386, "end": 1261.8193509615387, "probability": 0.9951171875 }, { "word": "注意", "start": 1261.8193509615387, "end": 1261.9993509615385, "probability": 0.9931640625 }, { "word": "力", "start": 1261.9993509615385, "end": 1262.1993509615386, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09403635540279733, "compression_ratio": 1.2727272727272727, "no_speech_prob": 5.203060204905796e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "完全没有做稠密注意力预热等训练量达到约34万亿tokens时他们会直接将上下文长度扩展到100万最终在总计45万亿tokens的多莫态数据上完成预训练因此CSAR CED和稀疏检索行为都不是后续才新增的模型实际上是从头学习这些机制的而如果一家实验室能够不依赖稠密注意力从零成功训练出稀疏注意力这通常就意味着这套注意力机制稳定可靠且具备很高的潜力后续" } ], "id": 583, "review_annotation": "", "row_id": 583, "bounded_row_id": 583, "parent_row_ids": [ 583 ], "parent_row_id": 583, "packet_id": 45, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1240.6593509615386, "packet_end": 1269.8473365384616, "raw_text": "而如果一家实验室能够不依赖稠密注意力", "selected_text": "而如果一家实验室能够不依赖稠密注意力", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 583 ], "packet_id": 45, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1262.1993509615386, "end": 1264.3793509615386, "text": "从零成功训练出稀疏注意力", "raw_asr_text": "从零成功训练出稀疏注意力", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 45, "words": [ { "word": "从", "start": 1262.1993509615386, "end": 1262.4593509615386, "probability": 0.9921875 }, { "word": "零", "start": 1262.4593509615386, "end": 1262.6793509615386, "probability": 0.9111328125 }, { "word": "成", "start": 1262.6793509615386, "end": 1262.8993509615386, "probability": 0.99609375 }, { "word": "功", "start": 1262.8993509615386, "end": 1263.0593509615387, "probability": 0.99951171875 }, { "word": "训", "start": 1263.0593509615387, "end": 1263.2993509615387, "probability": 0.9970703125 }, { "word": "练", "start": 1263.2993509615387, "end": 1263.4193509615386, "probability": 1.0 }, { "word": "出", "start": 1263.4193509615386, "end": 1263.5793509615387, "probability": 0.99169921875 }, { "word": "稀", "start": 1263.5793509615387, "end": 1263.8193509615387, "probability": 0.9833984375 }, { "word": "疏", "start": 1263.8193509615387, "end": 1263.9993509615385, "probability": 1.0 }, { "word": "注意", "start": 1263.9993509615385, "end": 1264.2393509615385, "probability": 0.99267578125 }, { "word": "力", "start": 1264.2393509615385, "end": 1264.3793509615386, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09403635540279733, "compression_ratio": 1.2727272727272727, "no_speech_prob": 5.203060204905796e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "完全没有做稠密注意力预热等训练量达到约34万亿tokens时他们会直接将上下文长度扩展到100万最终在总计45万亿tokens的多莫态数据上完成预训练因此CSAR CED和稀疏检索行为都不是后续才新增的模型实际上是从头学习这些机制的而如果一家实验室能够不依赖稠密注意力从零成功训练出稀疏注意力这通常就意味着这套注意力机制稳定可靠且具备很高的潜力后续" } ], "id": 584, "review_annotation": "", "row_id": 584, "bounded_row_id": 584, "parent_row_ids": [ 584 ], "parent_row_id": 584, "packet_id": 45, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1240.6593509615386, "packet_end": 1269.8473365384616, "raw_text": "从零成功训练出稀疏注意力", "selected_text": "从零成功训练出稀疏注意力", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 584 ], "packet_id": 45, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1264.3793509615386, "end": 1267.6793509615386, "text": "这通常就意味着这套注意力机制稳定可靠", "raw_asr_text": "这通常就意味着这套注意力机制稳定可靠", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 45, "words": [ { "word": "这", "start": 1264.3793509615386, "end": 1264.6993509615386, "probability": 0.9892578125 }, { "word": "通", "start": 1264.6993509615386, "end": 1264.9593509615386, "probability": 0.98974609375 }, { "word": "常", "start": 1264.9593509615386, "end": 1265.1193509615387, "probability": 0.99951171875 }, { "word": "就", "start": 1265.1193509615387, "end": 1265.2593509615385, "probability": 0.94775390625 }, { "word": "意", "start": 1265.2593509615385, "end": 1265.3993509615386, "probability": 0.99951171875 }, { "word": "味", "start": 1265.3993509615386, "end": 1265.4993509615385, "probability": 0.99853515625 }, { "word": "着", "start": 1265.4993509615385, "end": 1265.6193509615387, "probability": 0.99462890625 }, { "word": "这", "start": 1265.6193509615387, "end": 1265.7993509615387, "probability": 0.9970703125 }, { "word": "套", "start": 1265.7993509615387, "end": 1265.9593509615386, "probability": 0.998046875 }, { "word": "注意", "start": 1265.9593509615386, "end": 1266.1593509615386, "probability": 0.97412109375 }, { "word": "力", "start": 1266.1593509615386, "end": 1266.3193509615387, "probability": 0.998046875 }, { "word": "机", "start": 1266.3193509615387, "end": 1266.4993509615385, "probability": 0.99267578125 }, { "word": "制", "start": 1266.4993509615385, "end": 1266.6593509615386, "probability": 0.994140625 }, { "word": "稳", "start": 1266.6593509615386, "end": 1267.0393509615387, "probability": 0.994140625 }, { "word": "定", "start": 1267.0393509615387, "end": 1267.2393509615385, "probability": 0.99951171875 }, { "word": "可", "start": 1267.2393509615385, "end": 1267.4793509615386, "probability": 0.97607421875 }, { "word": "靠", "start": 1267.4793509615386, "end": 1267.6793509615386, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.09403635540279733, "compression_ratio": 1.2727272727272727, "no_speech_prob": 5.203060204905796e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "完全没有做稠密注意力预热等训练量达到约34万亿tokens时他们会直接将上下文长度扩展到100万最终在总计45万亿tokens的多莫态数据上完成预训练因此CSAR CED和稀疏检索行为都不是后续才新增的模型实际上是从头学习这些机制的而如果一家实验室能够不依赖稠密注意力从零成功训练出稀疏注意力这通常就意味着这套注意力机制稳定可靠且具备很高的潜力后续" } ], "id": 585, "review_annotation": "", "row_id": 585, "bounded_row_id": 585, "parent_row_ids": [ 585 ], "parent_row_id": 585, "packet_id": 45, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1240.6593509615386, "packet_end": 1269.8473365384616, "raw_text": "这通常就意味着这套注意力机制稳定可靠", "selected_text": "这通常就意味着这套注意力机制稳定可靠", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 585 ], "packet_id": 45, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1267.6793509615386, "end": 1269.0473365384616, "text": "且具备很高的潜力", "raw_asr_text": "且具备很高的潜力", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 45, "words": [ { "word": "且", "start": 1267.6793509615386, "end": 1268.0193509615385, "probability": 0.94970703125 }, { "word": "具", "start": 1268.0193509615385, "end": 1268.2193509615386, "probability": 1.0 }, { "word": "备", "start": 1268.2193509615386, "end": 1268.3793509615386, "probability": 0.9990234375 }, { "word": "很", "start": 1268.3793509615386, "end": 1268.6793509615386, "probability": 0.99658203125 }, { "word": "高", "start": 1268.6793509615386, "end": 1268.8793509615386, "probability": 0.99951171875 }, { "word": "的", "start": 1268.8793509615386, "end": 1268.9793509615386, "probability": 0.99853515625 }, { "word": "潜", "start": 1268.9793509615386, "end": 1269.1793509615386, "probability": 0.99658203125 }, { "word": "力", "start": 1269.1793509615386, "end": 1269.3593509615387, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.09403635540279733, "compression_ratio": 1.2727272727272727, "no_speech_prob": 5.203060204905796e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "完全没有做稠密注意力预热等训练量达到约34万亿tokens时他们会直接将上下文长度扩展到100万最终在总计45万亿tokens的多莫态数据上完成预训练因此CSAR CED和稀疏检索行为都不是后续才新增的模型实际上是从头学习这些机制的而如果一家实验室能够不依赖稠密注意力从零成功训练出稀疏注意力这通常就意味着这套注意力机制稳定可靠且具备很高的潜力后续" } ], "id": 586, "review_annotation": "", "row_id": 586, "bounded_row_id": 586, "parent_row_ids": [ 586 ], "parent_row_id": 586, "packet_id": 45, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1240.6593509615386, "packet_end": 1269.8473365384616, "raw_text": "且具备很高的潜力", "selected_text": "且具备很高的潜力", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 586 ], "packet_id": 45, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1269.3473365384616, "end": 1272.0273365384617, "text": "后训练部分的思路转变相当有意思", "raw_asr_text": "后训练部分的思路转变相当有意思", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 46, "words": [ { "word": "后", "start": 1269.3473365384616, "end": 1269.6873365384618, "probability": 0.6181640625 }, { "word": "训", "start": 1269.6873365384618, "end": 1269.8673365384616, "probability": 0.9931640625 }, { "word": "练", "start": 1269.8673365384616, "end": 1269.9873365384617, "probability": 0.998046875 }, { "word": "部分", "start": 1269.9873365384617, "end": 1270.2073365384617, "probability": 0.97607421875 }, { "word": "的", "start": 1270.2073365384617, "end": 1270.4273365384618, "probability": 0.98486328125 }, { "word": "思", "start": 1270.4273365384618, "end": 1270.5673365384616, "probability": 0.99755859375 }, { "word": "路", "start": 1270.5673365384616, "end": 1270.7273365384617, "probability": 0.99658203125 }, { "word": "转", "start": 1270.7273365384617, "end": 1270.9273365384618, "probability": 0.99609375 }, { "word": "变", "start": 1270.9273365384618, "end": 1271.1273365384618, "probability": 0.99853515625 }, { "word": "相", "start": 1271.1273365384618, "end": 1271.3873365384618, "probability": 0.99365234375 }, { "word": "当", "start": 1271.3873365384618, "end": 1271.5673365384616, "probability": 0.99853515625 }, { "word": "有", "start": 1271.5673365384616, "end": 1271.7673365384617, "probability": 0.99853515625 }, { "word": "意思", "start": 1271.7673365384617, "end": 1272.0273365384617, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.11620894762185904, "compression_ratio": 1.3588235294117648, "no_speech_prob": 4.6263864267315213e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "很高的潜力后训练部分的思路转变相当有意思因为Deep Seek明确表示他们这次基本没有推出新的后训练算法他们表示几乎所有提升都来自扩大模型的训练数据规模也就是靠数据扩容而非涉及新的强化学习目标我完全明白这是为什么Deep Seek已拥有最强且最稳定的后训练方案之一批词不变性让强化学习流程的可预测性大幅提升此外还有多教师在线策略蒸流看" } ], "id": 587, "review_annotation": "", "row_id": 587, "bounded_row_id": 587, "parent_row_ids": [ 587 ], "parent_row_id": 587, "packet_id": 46, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1268.2473365384617, "packet_end": 1297.4353221153845, "raw_text": "后训练部分的思路转变相当有意思", "selected_text": "后训练部分的思路转变相当有意思", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 587 ], "packet_id": 46, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1272.0273365384617, "end": 1273.7873365384617, "text": "因为Deep Seek明确表示", "raw_asr_text": "因为Deep Seek明确表示", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 46, "words": [ { "word": "因为", "start": 1272.0273365384617, "end": 1272.4473365384617, "probability": 0.9775390625 }, { "word": "De", "start": 1272.4473365384617, "end": 1272.6273365384618, "probability": 0.744140625 }, { "word": "ep", "start": 1272.6273365384618, "end": 1272.7273365384617, "probability": 0.97021484375 }, { "word": " Se", "start": 1272.7273365384617, "end": 1272.8673365384616, "probability": 0.4560546875 }, { "word": "ek", "start": 1272.8673365384616, "end": 1272.9873365384617, "probability": 0.98193359375 }, { "word": "明", "start": 1272.9873365384617, "end": 1273.2073365384617, "probability": 0.88671875 }, { "word": "确", "start": 1273.2073365384617, "end": 1273.4073365384618, "probability": 0.99951171875 }, { "word": "表示", "start": 1273.4073365384618, "end": 1273.7873365384617, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.11620894762185904, "compression_ratio": 1.3588235294117648, "no_speech_prob": 4.6263864267315213e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "很高的潜力后训练部分的思路转变相当有意思因为Deep Seek明确表示他们这次基本没有推出新的后训练算法他们表示几乎所有提升都来自扩大模型的训练数据规模也就是靠数据扩容而非涉及新的强化学习目标我完全明白这是为什么Deep Seek已拥有最强且最稳定的后训练方案之一批词不变性让强化学习流程的可预测性大幅提升此外还有多教师在线策略蒸流看" } ], "id": 588, "review_annotation": "", "row_id": 588, "bounded_row_id": 588, "parent_row_ids": [ 588 ], "parent_row_id": 588, "packet_id": 46, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1268.2473365384617, "packet_end": 1297.4353221153845, "raw_text": "因为Deep Seek明确表示", "selected_text": "因为Deep Seek明确表示", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 588 ], "packet_id": 46, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1273.7873365384617, "end": 1276.6073365384616, "text": "他们这次基本没有推出新的后训练算法", "raw_asr_text": "他们这次基本没有推出新的后训练算法", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 46, "words": [ { "word": "他们", "start": 1273.7873365384617, "end": 1274.1273365384618, "probability": 0.96435546875 }, { "word": "这", "start": 1274.1273365384618, "end": 1274.2873365384617, "probability": 0.998046875 }, { "word": "次", "start": 1274.2873365384617, "end": 1274.4673365384617, "probability": 0.98193359375 }, { "word": "基本", "start": 1274.4673365384617, "end": 1274.8073365384616, "probability": 0.994140625 }, { "word": "没有", "start": 1274.8073365384616, "end": 1275.0873365384616, "probability": 0.98876953125 }, { "word": "推", "start": 1275.0873365384616, "end": 1275.2473365384617, "probability": 0.99267578125 }, { "word": "出", "start": 1275.2473365384617, "end": 1275.4673365384617, "probability": 0.9990234375 }, { "word": "新", "start": 1275.4673365384617, "end": 1275.6473365384618, "probability": 0.998046875 }, { "word": "的", "start": 1275.6473365384618, "end": 1275.8073365384616, "probability": 0.99853515625 }, { "word": "后", "start": 1275.8073365384616, "end": 1275.9873365384617, "probability": 0.986328125 }, { "word": "训", "start": 1275.9873365384617, "end": 1276.1473365384618, "probability": 0.99951171875 }, { "word": "练", "start": 1276.1473365384618, "end": 1276.2473365384617, "probability": 0.9990234375 }, { "word": "算", "start": 1276.2473365384617, "end": 1276.4073365384618, "probability": 0.99365234375 }, { "word": "法", "start": 1276.4073365384618, "end": 1276.6073365384616, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11620894762185904, "compression_ratio": 1.3588235294117648, "no_speech_prob": 4.6263864267315213e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "很高的潜力后训练部分的思路转变相当有意思因为Deep Seek明确表示他们这次基本没有推出新的后训练算法他们表示几乎所有提升都来自扩大模型的训练数据规模也就是靠数据扩容而非涉及新的强化学习目标我完全明白这是为什么Deep Seek已拥有最强且最稳定的后训练方案之一批词不变性让强化学习流程的可预测性大幅提升此外还有多教师在线策略蒸流看" } ], "id": 589, "review_annotation": "", "row_id": 589, "bounded_row_id": 589, "parent_row_ids": [ 589 ], "parent_row_id": 589, "packet_id": 46, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1268.2473365384617, "packet_end": 1297.4353221153845, "raw_text": "他们这次基本没有推出新的后训练算法", "selected_text": "他们这次基本没有推出新的后训练算法", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 589 ], "packet_id": 46, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1276.6073365384616, "end": 1278.5673365384616, "text": "他们表示几乎所有提升", "raw_asr_text": "他们表示几乎所有提升", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 46, "words": [ { "word": "他们", "start": 1276.6073365384616, "end": 1277.1073365384616, "probability": 0.990234375 }, { "word": "表示", "start": 1277.1073365384616, "end": 1277.5073365384617, "probability": 0.99365234375 }, { "word": "几", "start": 1277.5073365384617, "end": 1277.8073365384616, "probability": 0.9775390625 }, { "word": "乎", "start": 1277.8073365384616, "end": 1277.9073365384618, "probability": 1.0 }, { "word": "所有", "start": 1277.9073365384618, "end": 1278.2273365384617, "probability": 0.99853515625 }, { "word": "提", "start": 1278.2273365384617, "end": 1278.4073365384618, "probability": 0.99609375 }, { "word": "升", "start": 1278.4073365384618, "end": 1278.5673365384616, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.11620894762185904, "compression_ratio": 1.3588235294117648, "no_speech_prob": 4.6263864267315213e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "很高的潜力后训练部分的思路转变相当有意思因为Deep Seek明确表示他们这次基本没有推出新的后训练算法他们表示几乎所有提升都来自扩大模型的训练数据规模也就是靠数据扩容而非涉及新的强化学习目标我完全明白这是为什么Deep Seek已拥有最强且最稳定的后训练方案之一批词不变性让强化学习流程的可预测性大幅提升此外还有多教师在线策略蒸流看" } ], "id": 590, "review_annotation": "", "row_id": 590, "bounded_row_id": 590, "parent_row_ids": [ 590 ], "parent_row_id": 590, "packet_id": 46, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1268.2473365384617, "packet_end": 1297.4353221153845, "raw_text": "他们表示几乎所有提升", "selected_text": "他们表示几乎所有提升", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 590 ], "packet_id": 46, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1278.5673365384616, "end": 1281.1873365384618, "text": "都来自扩大模型的训练数据规模", "raw_asr_text": "都来自扩大模型的训练数据规模", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 46, "words": [ { "word": "都", "start": 1278.5673365384616, "end": 1278.7673365384617, "probability": 0.98046875 }, { "word": "来", "start": 1278.7673365384617, "end": 1278.9473365384617, "probability": 0.99853515625 }, { "word": "自", "start": 1278.9473365384617, "end": 1279.1073365384616, "probability": 0.9990234375 }, { "word": "扩", "start": 1279.1073365384616, "end": 1279.4473365384617, "probability": 0.9853515625 }, { "word": "大", "start": 1279.4473365384617, "end": 1279.6473365384618, "probability": 0.99853515625 }, { "word": "模", "start": 1279.6473365384618, "end": 1279.9073365384618, "probability": 0.99609375 }, { "word": "型", "start": 1279.9073365384618, "end": 1280.1073365384616, "probability": 0.998046875 }, { "word": "的", "start": 1280.1073365384616, "end": 1280.2473365384617, "probability": 0.99853515625 }, { "word": "训", "start": 1280.2473365384617, "end": 1280.4673365384617, "probability": 0.9970703125 }, { "word": "练", "start": 1280.4673365384617, "end": 1280.5873365384616, "probability": 0.99951171875 }, { "word": "数", "start": 1280.5873365384616, "end": 1280.7273365384617, "probability": 0.99755859375 }, { "word": "据", "start": 1280.7273365384617, "end": 1280.8073365384616, "probability": 0.9990234375 }, { "word": "规", "start": 1280.8073365384616, "end": 1280.9673365384617, "probability": 0.9990234375 }, { "word": "模", "start": 1280.9673365384617, "end": 1281.1873365384618, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11620894762185904, "compression_ratio": 1.3588235294117648, "no_speech_prob": 4.6263864267315213e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "很高的潜力后训练部分的思路转变相当有意思因为Deep Seek明确表示他们这次基本没有推出新的后训练算法他们表示几乎所有提升都来自扩大模型的训练数据规模也就是靠数据扩容而非涉及新的强化学习目标我完全明白这是为什么Deep Seek已拥有最强且最稳定的后训练方案之一批词不变性让强化学习流程的可预测性大幅提升此外还有多教师在线策略蒸流看" } ], "id": 591, "review_annotation": "", "row_id": 591, "bounded_row_id": 591, "parent_row_ids": [ 591 ], "parent_row_id": 591, "packet_id": 46, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1268.2473365384617, "packet_end": 1297.4353221153845, "raw_text": "都来自扩大模型的训练数据规模", "selected_text": "都来自扩大模型的训练数据规模", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 591 ], "packet_id": 46, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1281.1873365384618, "end": 1282.8073365384616, "text": "也就是靠数据扩容", "raw_asr_text": "也就是靠数据扩容", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 46, "words": [ { "word": "也", "start": 1281.1873365384618, "end": 1281.5273365384617, "probability": 0.9580078125 }, { "word": "就是", "start": 1281.5273365384617, "end": 1281.7873365384617, "probability": 0.99658203125 }, { "word": "靠", "start": 1281.7873365384617, "end": 1282.0273365384617, "probability": 0.99609375 }, { "word": "数", "start": 1282.0273365384617, "end": 1282.2873365384617, "probability": 0.9970703125 }, { "word": "据", "start": 1282.2873365384617, "end": 1282.4073365384618, "probability": 1.0 }, { "word": "扩", "start": 1282.4073365384618, "end": 1282.5873365384616, "probability": 0.99609375 }, { "word": "容", "start": 1282.5873365384616, "end": 1282.8073365384616, "probability": 0.98876953125 } ], "asr_metrics": { "avg_logprob": -0.11620894762185904, "compression_ratio": 1.3588235294117648, "no_speech_prob": 4.6263864267315213e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "很高的潜力后训练部分的思路转变相当有意思因为Deep Seek明确表示他们这次基本没有推出新的后训练算法他们表示几乎所有提升都来自扩大模型的训练数据规模也就是靠数据扩容而非涉及新的强化学习目标我完全明白这是为什么Deep Seek已拥有最强且最稳定的后训练方案之一批词不变性让强化学习流程的可预测性大幅提升此外还有多教师在线策略蒸流看" } ], "id": 592, "review_annotation": "", "row_id": 592, "bounded_row_id": 592, "parent_row_ids": [ 592 ], "parent_row_id": 592, "packet_id": 46, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1268.2473365384617, "packet_end": 1297.4353221153845, "raw_text": "也就是靠数据扩容", "selected_text": "也就是靠数据扩容", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 592 ], "packet_id": 46, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1282.8073365384616, "end": 1285.1473365384618, "text": "而非涉及新的强化学习目标", "raw_asr_text": "而非涉及新的强化学习目标", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 46, "words": [ { "word": "而", "start": 1282.8073365384616, "end": 1283.0673365384616, "probability": 0.978515625 }, { "word": "非", "start": 1283.0673365384616, "end": 1283.2273365384617, "probability": 0.998046875 }, { "word": "涉", "start": 1283.2273365384617, "end": 1283.4273365384618, "probability": 0.94580078125 }, { "word": "及", "start": 1283.4273365384618, "end": 1283.5673365384616, "probability": 0.99951171875 }, { "word": "新", "start": 1283.5673365384616, "end": 1283.8673365384616, "probability": 0.98974609375 }, { "word": "的", "start": 1283.8673365384616, "end": 1284.0473365384616, "probability": 0.9951171875 }, { "word": "强", "start": 1284.0473365384616, "end": 1284.3673365384616, "probability": 0.99560546875 }, { "word": "化", "start": 1284.3673365384616, "end": 1284.5673365384616, "probability": 0.99951171875 }, { "word": "学", "start": 1284.5673365384616, "end": 1284.7273365384617, "probability": 0.99951171875 }, { "word": "习", "start": 1284.7273365384617, "end": 1284.8473365384616, "probability": 1.0 }, { "word": "目", "start": 1284.8473365384616, "end": 1284.9873365384617, "probability": 0.99853515625 }, { "word": "标", "start": 1284.9873365384617, "end": 1285.1473365384618, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11620894762185904, "compression_ratio": 1.3588235294117648, "no_speech_prob": 4.6263864267315213e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "很高的潜力后训练部分的思路转变相当有意思因为Deep Seek明确表示他们这次基本没有推出新的后训练算法他们表示几乎所有提升都来自扩大模型的训练数据规模也就是靠数据扩容而非涉及新的强化学习目标我完全明白这是为什么Deep Seek已拥有最强且最稳定的后训练方案之一批词不变性让强化学习流程的可预测性大幅提升此外还有多教师在线策略蒸流看" } ], "id": 593, "review_annotation": "", "row_id": 593, "bounded_row_id": 593, "parent_row_ids": [ 593 ], "parent_row_id": 593, "packet_id": 46, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1268.2473365384617, "packet_end": 1297.4353221153845, "raw_text": "而非涉及新的强化学习目标", "selected_text": "而非涉及新的强化学习目标", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 593 ], "packet_id": 46, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1285.1473365384618, "end": 1286.9273365384618, "text": "我完全明白这是为什么", "raw_asr_text": "我完全明白这是为什么", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 46, "words": [ { "word": "我", "start": 1285.1473365384618, "end": 1285.4073365384618, "probability": 0.99169921875 }, { "word": "完全", "start": 1285.4073365384618, "end": 1285.7673365384617, "probability": 0.99853515625 }, { "word": "明白", "start": 1285.7673365384617, "end": 1286.1273365384618, "probability": 0.99755859375 }, { "word": "这是", "start": 1286.1273365384618, "end": 1286.4273365384618, "probability": 0.9521484375 }, { "word": "为什么", "start": 1286.4273365384618, "end": 1286.9273365384618, "probability": 0.990234375 } ], "asr_metrics": { "avg_logprob": -0.11620894762185904, "compression_ratio": 1.3588235294117648, "no_speech_prob": 4.6263864267315213e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "很高的潜力后训练部分的思路转变相当有意思因为Deep Seek明确表示他们这次基本没有推出新的后训练算法他们表示几乎所有提升都来自扩大模型的训练数据规模也就是靠数据扩容而非涉及新的强化学习目标我完全明白这是为什么Deep Seek已拥有最强且最稳定的后训练方案之一批词不变性让强化学习流程的可预测性大幅提升此外还有多教师在线策略蒸流看" } ], "id": 594, "review_annotation": "", "row_id": 594, "bounded_row_id": 594, "parent_row_ids": [ 594 ], "parent_row_id": 594, "packet_id": 46, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1268.2473365384617, "packet_end": 1297.4353221153845, "raw_text": "我完全明白这是为什么", "selected_text": "我完全明白这是为什么", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 594 ], "packet_id": 46, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1286.9273365384618, "end": 1288.6873365384618, "text": "Deep Seek已拥有最强", "raw_asr_text": "Deep Seek已拥有最强", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 46, "words": [ { "word": "De", "start": 1286.9273365384618, "end": 1287.3073365384616, "probability": 0.97607421875 }, { "word": "ep", "start": 1287.3073365384616, "end": 1287.4473365384617, "probability": 0.99951171875 }, { "word": " Se", "start": 1287.4473365384617, "end": 1287.5873365384616, "probability": 0.98046875 }, { "word": "ek", "start": 1287.5873365384616, "end": 1287.7073365384617, "probability": 1.0 }, { "word": "已", "start": 1287.7073365384617, "end": 1287.9673365384617, "probability": 0.93115234375 }, { "word": "拥", "start": 1287.9673365384617, "end": 1288.1073365384616, "probability": 0.99951171875 }, { "word": "有", "start": 1288.1073365384616, "end": 1288.2473365384617, "probability": 0.99951171875 }, { "word": "最", "start": 1288.2473365384617, "end": 1288.4473365384617, "probability": 0.98681640625 }, { "word": "强", "start": 1288.4473365384617, "end": 1288.6873365384618, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.11620894762185904, "compression_ratio": 1.3588235294117648, "no_speech_prob": 4.6263864267315213e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "很高的潜力后训练部分的思路转变相当有意思因为Deep Seek明确表示他们这次基本没有推出新的后训练算法他们表示几乎所有提升都来自扩大模型的训练数据规模也就是靠数据扩容而非涉及新的强化学习目标我完全明白这是为什么Deep Seek已拥有最强且最稳定的后训练方案之一批词不变性让强化学习流程的可预测性大幅提升此外还有多教师在线策略蒸流看" } ], "id": 595, "review_annotation": "", "row_id": 595, "bounded_row_id": 595, "parent_row_ids": [ 595 ], "parent_row_id": 595, "packet_id": 46, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1268.2473365384617, "packet_end": 1297.4353221153845, "raw_text": "Deep Seek已拥有最强", "selected_text": "Deep Seek已拥有最强", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 595 ], "packet_id": 46, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1288.6873365384618, "end": 1290.8473365384616, "text": "且最稳定的后训练方案之一", "raw_asr_text": "且最稳定的后训练方案之一", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 46, "words": [ { "word": "且", "start": 1288.6873365384618, "end": 1289.2073365384617, "probability": 0.98681640625 }, { "word": "最", "start": 1289.2073365384617, "end": 1289.4473365384617, "probability": 0.99169921875 }, { "word": "稳", "start": 1289.4473365384617, "end": 1289.6473365384618, "probability": 0.998046875 }, { "word": "定", "start": 1289.6473365384618, "end": 1289.7673365384617, "probability": 0.99951171875 }, { "word": "的", "start": 1289.7673365384617, "end": 1289.8873365384618, "probability": 0.998046875 }, { "word": "后", "start": 1289.8873365384618, "end": 1290.0073365384617, "probability": 0.98681640625 }, { "word": "训", "start": 1290.0073365384617, "end": 1290.1473365384618, "probability": 0.9990234375 }, { "word": "练", "start": 1290.1473365384618, "end": 1290.2673365384617, "probability": 0.998046875 }, { "word": "方", "start": 1290.2673365384617, "end": 1290.3873365384618, "probability": 0.9990234375 }, { "word": "案", "start": 1290.3873365384618, "end": 1290.5273365384617, "probability": 0.998046875 }, { "word": "之", "start": 1290.5273365384617, "end": 1290.6673365384618, "probability": 0.998046875 }, { "word": "一", "start": 1290.6673365384618, "end": 1290.8473365384616, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.11620894762185904, "compression_ratio": 1.3588235294117648, "no_speech_prob": 4.6263864267315213e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "很高的潜力后训练部分的思路转变相当有意思因为Deep Seek明确表示他们这次基本没有推出新的后训练算法他们表示几乎所有提升都来自扩大模型的训练数据规模也就是靠数据扩容而非涉及新的强化学习目标我完全明白这是为什么Deep Seek已拥有最强且最稳定的后训练方案之一批词不变性让强化学习流程的可预测性大幅提升此外还有多教师在线策略蒸流看" } ], "id": 596, "review_annotation": "", "row_id": 596, "bounded_row_id": 596, "parent_row_ids": [ 596 ], "parent_row_id": 596, "packet_id": 46, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1268.2473365384617, "packet_end": 1297.4353221153845, "raw_text": "且最稳定的后训练方案之一", "selected_text": "且最稳定的后训练方案之一", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 596 ], "packet_id": 46, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1290.8473365384616, "end": 1291.7073365384617, "text": "批词不变性", "raw_asr_text": "批词不变性", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 46, "words": [ { "word": "批", "start": 1290.8473365384616, "end": 1291.0873365384616, "probability": 0.99462890625 }, { "word": "词", "start": 1291.0873365384616, "end": 1291.2673365384617, "probability": 0.830078125 }, { "word": "不", "start": 1291.2673365384617, "end": 1291.3873365384618, "probability": 0.9951171875 }, { "word": "变", "start": 1291.3873365384618, "end": 1291.5273365384617, "probability": 0.982421875 }, { "word": "性", "start": 1291.5273365384617, "end": 1291.7073365384617, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.11620894762185904, "compression_ratio": 1.3588235294117648, "no_speech_prob": 4.6263864267315213e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "很高的潜力后训练部分的思路转变相当有意思因为Deep Seek明确表示他们这次基本没有推出新的后训练算法他们表示几乎所有提升都来自扩大模型的训练数据规模也就是靠数据扩容而非涉及新的强化学习目标我完全明白这是为什么Deep Seek已拥有最强且最稳定的后训练方案之一批词不变性让强化学习流程的可预测性大幅提升此外还有多教师在线策略蒸流看" } ], "id": 597, "review_annotation": "", "row_id": 597, "bounded_row_id": 597, "parent_row_ids": [ 597 ], "parent_row_id": 597, "packet_id": 46, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1268.2473365384617, "packet_end": 1297.4353221153845, "raw_text": "批词不变性", "selected_text": "批词不变性", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 597 ], "packet_id": 46, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1291.7073365384617, "end": 1294.4873365384617, "text": "让强化学习流程的可预测性大幅提升", "raw_asr_text": "让强化学习流程的可预测性大幅提升", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 46, "words": [ { "word": "让", "start": 1291.7073365384617, "end": 1291.9673365384617, "probability": 0.97314453125 }, { "word": "强", "start": 1291.9673365384617, "end": 1292.2673365384617, "probability": 0.9990234375 }, { "word": "化", "start": 1292.2673365384617, "end": 1292.4673365384617, "probability": 0.99951171875 }, { "word": "学", "start": 1292.4673365384617, "end": 1292.6073365384616, "probability": 0.998046875 }, { "word": "习", "start": 1292.6073365384616, "end": 1292.7273365384617, "probability": 1.0 }, { "word": "流", "start": 1292.7273365384617, "end": 1292.8673365384616, "probability": 0.99560546875 }, { "word": "程", "start": 1292.8673365384616, "end": 1292.9873365384617, "probability": 0.99951171875 }, { "word": "的", "start": 1292.9873365384617, "end": 1293.0873365384616, "probability": 0.99462890625 }, { "word": "可", "start": 1293.0873365384616, "end": 1293.2673365384617, "probability": 0.99755859375 }, { "word": "预", "start": 1293.2673365384617, "end": 1293.3873365384618, "probability": 0.9990234375 }, { "word": "测", "start": 1293.3873365384618, "end": 1293.5673365384616, "probability": 0.9990234375 }, { "word": "性", "start": 1293.5673365384616, "end": 1293.7073365384617, "probability": 0.998046875 }, { "word": "大", "start": 1293.7073365384617, "end": 1293.9673365384617, "probability": 0.99609375 }, { "word": "幅", "start": 1293.9673365384617, "end": 1294.1673365384618, "probability": 0.99951171875 }, { "word": "提", "start": 1294.1673365384618, "end": 1294.2873365384617, "probability": 0.99755859375 }, { "word": "升", "start": 1294.2873365384617, "end": 1294.4873365384617, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.11620894762185904, "compression_ratio": 1.3588235294117648, "no_speech_prob": 4.6263864267315213e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "很高的潜力后训练部分的思路转变相当有意思因为Deep Seek明确表示他们这次基本没有推出新的后训练算法他们表示几乎所有提升都来自扩大模型的训练数据规模也就是靠数据扩容而非涉及新的强化学习目标我完全明白这是为什么Deep Seek已拥有最强且最稳定的后训练方案之一批词不变性让强化学习流程的可预测性大幅提升此外还有多教师在线策略蒸流看" } ], "id": 598, "review_annotation": "", "row_id": 598, "bounded_row_id": 598, "parent_row_ids": [ 598 ], "parent_row_id": 598, "packet_id": 46, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1268.2473365384617, "packet_end": 1297.4353221153845, "raw_text": "让强化学习流程的可预测性大幅提升", "selected_text": "让强化学习流程的可预测性大幅提升", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 598 ], "packet_id": 46, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1294.4873365384617, "end": 1296.6353221153845, "text": "此外还有多教师在线策略蒸流", "raw_asr_text": "此外还有多教师在线策略蒸流", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 46, "words": [ { "word": "此", "start": 1294.4873365384617, "end": 1294.7873365384617, "probability": 0.99658203125 }, { "word": "外", "start": 1294.7873365384617, "end": 1294.9673365384617, "probability": 0.9990234375 }, { "word": "还有", "start": 1294.9673365384617, "end": 1295.2073365384617, "probability": 0.96435546875 }, { "word": "多", "start": 1295.2073365384617, "end": 1295.4473365384617, "probability": 0.97802734375 }, { "word": "教", "start": 1295.4473365384617, "end": 1295.7073365384617, "probability": 0.99560546875 }, { "word": "师", "start": 1295.7073365384617, "end": 1295.9273365384618, "probability": 0.998046875 }, { "word": "在", "start": 1295.9273365384618, "end": 1296.1073365384616, "probability": 0.9697265625 }, { "word": "线", "start": 1296.1073365384616, "end": 1296.3873365384618, "probability": 0.99755859375 }, { "word": "策", "start": 1296.3873365384618, "end": 1296.6673365384618, "probability": 0.9951171875 }, { "word": "略", "start": 1296.6673365384618, "end": 1296.8473365384616, "probability": 1.0 }, { "word": "蒸", "start": 1296.8473365384616, "end": 1296.9873365384617, "probability": 0.66796875 }, { "word": "流", "start": 1296.9873365384617, "end": 1297.1873365384618, "probability": 0.2900390625 } ], "asr_metrics": { "avg_logprob": -0.11620894762185904, "compression_ratio": 1.3588235294117648, "no_speech_prob": 4.6263864267315213e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "很高的潜力后训练部分的思路转变相当有意思因为Deep Seek明确表示他们这次基本没有推出新的后训练算法他们表示几乎所有提升都来自扩大模型的训练数据规模也就是靠数据扩容而非涉及新的强化学习目标我完全明白这是为什么Deep Seek已拥有最强且最稳定的后训练方案之一批词不变性让强化学习流程的可预测性大幅提升此外还有多教师在线策略蒸流看" } ], "id": 599, "review_annotation": "", "row_id": 599, "bounded_row_id": 599, "parent_row_ids": [ 599 ], "parent_row_id": 599, "packet_id": 46, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1268.2473365384617, "packet_end": 1297.4353221153845, "raw_text": "此外还有多教师在线策略蒸流", "selected_text": "此外还有多教师在线策略蒸流", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 599 ], "packet_id": 46, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1297.1953221153844, "end": 1298.4153221153845, "text": "他们现在将其称为", "raw_asr_text": "他们现在将其称为", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "他们", "start": 1297.1953221153844, "end": 1297.4553221153844, "probability": 0.64453125 }, { "word": "现在", "start": 1297.4553221153844, "end": 1297.7353221153846, "probability": 0.9921875 }, { "word": "将", "start": 1297.7353221153846, "end": 1297.9753221153846, "probability": 0.998046875 }, { "word": "其", "start": 1297.9753221153846, "end": 1298.1153221153845, "probability": 0.99951171875 }, { "word": "称", "start": 1298.1153221153845, "end": 1298.2553221153846, "probability": 0.990234375 }, { "word": "为", "start": 1298.2553221153846, "end": 1298.4153221153845, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 600, "review_annotation": "", "row_id": 600, "bounded_row_id": 600, "parent_row_ids": [ 600 ], "parent_row_id": 600, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "他们现在将其称为", "selected_text": "他们现在将其称为", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 600 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1298.4153221153845, "end": 1300.0553221153846, "text": "大规模同策略蒸流", "raw_asr_text": "大规模同策略蒸流", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "大", "start": 1298.4153221153845, "end": 1298.6353221153845, "probability": 0.97607421875 }, { "word": "规", "start": 1298.6353221153845, "end": 1298.8353221153845, "probability": 1.0 }, { "word": "模", "start": 1298.8353221153845, "end": 1298.9953221153846, "probability": 0.99951171875 }, { "word": "同", "start": 1298.9953221153846, "end": 1299.2753221153846, "probability": 0.9501953125 }, { "word": "策", "start": 1299.2753221153846, "end": 1299.5353221153846, "probability": 0.998046875 }, { "word": "略", "start": 1299.5353221153846, "end": 1299.6953221153844, "probability": 1.0 }, { "word": "蒸", "start": 1299.6953221153844, "end": 1299.8753221153845, "probability": 0.99951171875 }, { "word": "流", "start": 1299.8753221153845, "end": 1300.0553221153846, "probability": 0.9873046875 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 601, "review_annotation": "", "row_id": 601, "bounded_row_id": 601, "parent_row_ids": [ 601 ], "parent_row_id": 601, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "大规模同策略蒸流", "selected_text": "大规模同策略蒸流", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 601 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1300.0553221153846, "end": 1302.0953221153845, "text": "V4.1会向不同领域的", "raw_asr_text": "V4.1会向不同领域的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "V", "start": 1300.0553221153846, "end": 1300.3353221153845, "probability": 0.8818359375 }, { "word": "4.", "start": 1300.3353221153845, "end": 1300.6153221153845, "probability": 0.9609375 }, { "word": "1", "start": 1300.7553221153846, "end": 1300.9553221153844, "probability": 0.99609375 }, { "word": "会", "start": 1300.9553221153844, "end": 1301.1553221153845, "probability": 0.9501953125 }, { "word": "向", "start": 1301.1553221153845, "end": 1301.3353221153845, "probability": 0.99609375 }, { "word": "不同", "start": 1301.3353221153845, "end": 1301.6153221153845, "probability": 0.99853515625 }, { "word": "领", "start": 1301.6153221153845, "end": 1301.8153221153846, "probability": 1.0 }, { "word": "域", "start": 1301.8153221153846, "end": 1301.9553221153844, "probability": 1.0 }, { "word": "的", "start": 1301.9553221153844, "end": 1302.0953221153845, "probability": 0.99609375 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 602, "review_annotation": "", "row_id": 602, "bounded_row_id": 602, "parent_row_ids": [ 602 ], "parent_row_id": 602, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "V4.1会向不同领域的", "selected_text": "V4.1会向不同领域的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 602 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1302.0953221153845, "end": 1304.7353221153846, "text": "40多个各有专攻的教师模型学习", "raw_asr_text": "40多个各有专攻的教师模型学习", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "40", "start": 1302.0953221153845, "end": 1302.4153221153845, "probability": 0.97265625 }, { "word": "多", "start": 1302.4153221153845, "end": 1302.6153221153845, "probability": 0.99462890625 }, { "word": "个", "start": 1302.6153221153845, "end": 1302.7953221153846, "probability": 0.9990234375 }, { "word": "各", "start": 1302.7953221153846, "end": 1303.0353221153846, "probability": 0.9638671875 }, { "word": "有", "start": 1303.0353221153846, "end": 1303.2553221153846, "probability": 0.99951171875 }, { "word": "专", "start": 1303.2553221153846, "end": 1303.4953221153846, "probability": 0.9990234375 }, { "word": "攻", "start": 1303.4953221153846, "end": 1303.6153221153845, "probability": 0.80419921875 }, { "word": "的", "start": 1303.6153221153845, "end": 1303.7353221153846, "probability": 0.9990234375 }, { "word": "教", "start": 1303.7353221153846, "end": 1303.8753221153845, "probability": 0.99853515625 }, { "word": "师", "start": 1303.8753221153845, "end": 1303.9953221153846, "probability": 0.8369140625 }, { "word": "模", "start": 1303.9953221153846, "end": 1304.1353221153845, "probability": 1.0 }, { "word": "型", "start": 1304.1353221153845, "end": 1304.3153221153846, "probability": 0.99951171875 }, { "word": "学", "start": 1304.3153221153846, "end": 1304.4753221153846, "probability": 0.9990234375 }, { "word": "习", "start": 1304.4753221153846, "end": 1304.7353221153846, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 603, "review_annotation": "", "row_id": 603, "bounded_row_id": 603, "parent_row_ids": [ 603 ], "parent_row_id": 603, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "40多个各有专攻的教师模型学习", "selected_text": "40多个各有专攻的教师模型学习", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 603 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1304.7353221153846, "end": 1306.6953221153844, "text": "这也说明DeepSeek目前认为", "raw_asr_text": "这也说明DeepSeek目前认为", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "这", "start": 1304.7353221153846, "end": 1305.0353221153846, "probability": 0.9951171875 }, { "word": "也", "start": 1305.0353221153846, "end": 1305.1353221153845, "probability": 0.99755859375 }, { "word": "说", "start": 1305.1353221153845, "end": 1305.2553221153846, "probability": 0.99951171875 }, { "word": "明", "start": 1305.2553221153846, "end": 1305.4753221153846, "probability": 0.99951171875 }, { "word": "De", "start": 1305.4753221153846, "end": 1305.6553221153845, "probability": 0.869140625 }, { "word": "ep", "start": 1305.6553221153845, "end": 1305.7353221153846, "probability": 0.94580078125 }, { "word": "Se", "start": 1305.7353221153846, "end": 1305.8753221153845, "probability": 0.46484375 }, { "word": "ek", "start": 1305.8753221153845, "end": 1305.9953221153846, "probability": 0.73828125 }, { "word": "目前", "start": 1305.9953221153846, "end": 1306.2953221153846, "probability": 0.9921875 }, { "word": "认", "start": 1306.2953221153846, "end": 1306.5353221153846, "probability": 0.998046875 }, { "word": "为", "start": 1306.5353221153846, "end": 1306.6953221153844, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 604, "review_annotation": "", "row_id": 604, "bounded_row_id": 604, "parent_row_ids": [ 604 ], "parent_row_id": 604, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "这也说明DeepSeek目前认为", "selected_text": "这也说明DeepSeek目前认为", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 604 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1306.6953221153844, "end": 1308.9953221153846, "text": "投入更多时间和资金优化数据", "raw_asr_text": "投入更多时间和资金优化数据", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "投", "start": 1306.6953221153844, "end": 1306.9353221153845, "probability": 0.9970703125 }, { "word": "入", "start": 1306.9353221153845, "end": 1307.0953221153845, "probability": 0.99951171875 }, { "word": "更", "start": 1307.0953221153845, "end": 1307.2753221153846, "probability": 0.99951171875 }, { "word": "多", "start": 1307.2753221153846, "end": 1307.4353221153845, "probability": 1.0 }, { "word": "时间", "start": 1307.4353221153845, "end": 1307.7753221153846, "probability": 0.998046875 }, { "word": "和", "start": 1307.7753221153846, "end": 1307.9953221153846, "probability": 0.9990234375 }, { "word": "资", "start": 1307.9953221153846, "end": 1308.1553221153845, "probability": 1.0 }, { "word": "金", "start": 1308.1553221153845, "end": 1308.2953221153846, "probability": 0.99853515625 }, { "word": "优", "start": 1308.2953221153846, "end": 1308.4353221153845, "probability": 0.9990234375 }, { "word": "化", "start": 1308.4353221153845, "end": 1308.5753221153846, "probability": 0.99951171875 }, { "word": "数", "start": 1308.5753221153846, "end": 1308.7553221153846, "probability": 0.9990234375 }, { "word": "据", "start": 1308.7553221153846, "end": 1308.9953221153846, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 605, "review_annotation": "", "row_id": 605, "bounded_row_id": 605, "parent_row_ids": [ 605 ], "parent_row_id": 605, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "投入更多时间和资金优化数据", "selected_text": "投入更多时间和资金优化数据", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 605 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1308.9953221153846, "end": 1311.2753221153846, "text": "比继续研发新的后训练算法", "raw_asr_text": "比继续研发新的后训练算法", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "比", "start": 1308.9953221153846, "end": 1309.4753221153846, "probability": 0.99755859375 }, { "word": "继", "start": 1309.4753221153846, "end": 1309.7353221153846, "probability": 0.998046875 }, { "word": "续", "start": 1309.7353221153846, "end": 1309.8353221153845, "probability": 1.0 }, { "word": "研", "start": 1309.8353221153845, "end": 1309.9953221153846, "probability": 0.9990234375 }, { "word": "发", "start": 1309.9953221153846, "end": 1310.1753221153845, "probability": 0.9990234375 }, { "word": "新", "start": 1310.1753221153845, "end": 1310.3953221153845, "probability": 0.99853515625 }, { "word": "的", "start": 1310.3953221153845, "end": 1310.5353221153846, "probability": 1.0 }, { "word": "后", "start": 1310.5353221153846, "end": 1310.6553221153845, "probability": 0.9951171875 }, { "word": "训", "start": 1310.6553221153845, "end": 1310.7953221153846, "probability": 0.99560546875 }, { "word": "练", "start": 1310.7953221153846, "end": 1310.9153221153845, "probability": 0.9990234375 }, { "word": "算", "start": 1310.9153221153845, "end": 1311.0553221153846, "probability": 0.99658203125 }, { "word": "法", "start": 1311.0553221153846, "end": 1311.2753221153846, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 606, "review_annotation": "", "row_id": 606, "bounded_row_id": 606, "parent_row_ids": [ 606 ], "parent_row_id": 606, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "比继续研发新的后训练算法", "selected_text": "比继续研发新的后训练算法", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 606 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1311.2753221153846, "end": 1312.0953221153845, "text": "更有价值", "raw_asr_text": "更有价值", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "更", "start": 1311.2753221153846, "end": 1311.4953221153846, "probability": 0.99853515625 }, { "word": "有", "start": 1311.4953221153846, "end": 1311.6553221153845, "probability": 0.99951171875 }, { "word": "价", "start": 1311.6553221153845, "end": 1311.8753221153845, "probability": 0.99951171875 }, { "word": "值", "start": 1311.8753221153845, "end": 1312.0953221153845, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 607, "review_annotation": "", "row_id": 607, "bounded_row_id": 607, "parent_row_ids": [ 607 ], "parent_row_id": 607, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "更有价值", "selected_text": "更有价值", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 607 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1312.0953221153845, "end": 1313.6953221153844, "text": "尤其是本篇论文中", "raw_asr_text": "尤其是本篇论文中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "尤", "start": 1312.0953221153845, "end": 1312.6553221153845, "probability": 0.9892578125 }, { "word": "其", "start": 1312.6553221153845, "end": 1312.7553221153846, "probability": 1.0 }, { "word": "是", "start": 1312.7553221153846, "end": 1312.8753221153845, "probability": 1.0 }, { "word": "本", "start": 1312.8753221153845, "end": 1313.0553221153846, "probability": 0.9990234375 }, { "word": "篇", "start": 1313.0553221153846, "end": 1313.2153221153847, "probability": 0.99609375 }, { "word": "论", "start": 1313.2153221153847, "end": 1313.3353221153845, "probability": 0.9990234375 }, { "word": "文", "start": 1313.3353221153845, "end": 1313.5153221153846, "probability": 0.99951171875 }, { "word": "中", "start": 1313.5153221153846, "end": 1313.6953221153844, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 608, "review_annotation": "", "row_id": 608, "bounded_row_id": 608, "parent_row_ids": [ 608 ], "parent_row_id": 608, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "尤其是本篇论文中", "selected_text": "尤其是本篇论文中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 608 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1313.6953221153844, "end": 1314.7553221153846, "text": "他们十分重视", "raw_asr_text": "他们十分重视", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "他们", "start": 1313.6953221153844, "end": 1313.9553221153844, "probability": 0.98681640625 }, { "word": "十", "start": 1313.9553221153844, "end": 1314.2353221153846, "probability": 0.990234375 }, { "word": "分", "start": 1314.2353221153846, "end": 1314.4153221153845, "probability": 0.99951171875 }, { "word": "重", "start": 1314.4153221153845, "end": 1314.5153221153846, "probability": 0.9990234375 }, { "word": "视", "start": 1314.5153221153846, "end": 1314.7553221153846, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 609, "review_annotation": "", "row_id": 609, "bounded_row_id": 609, "parent_row_ids": [ 609 ], "parent_row_id": 609, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "他们十分重视", "selected_text": "他们十分重视", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 609 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1314.7553221153846, "end": 1315.8753221153845, "text": "强化学习数据的", "raw_asr_text": "强化学习数据的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "强", "start": 1314.7553221153846, "end": 1314.9953221153846, "probability": 0.9990234375 }, { "word": "化", "start": 1314.9953221153846, "end": 1315.1553221153845, "probability": 0.99853515625 }, { "word": "学", "start": 1315.1553221153845, "end": 1315.3153221153846, "probability": 1.0 }, { "word": "习", "start": 1315.3153221153846, "end": 1315.4353221153845, "probability": 0.9990234375 }, { "word": "数", "start": 1315.4353221153845, "end": 1315.6153221153845, "probability": 0.9990234375 }, { "word": "据", "start": 1315.6153221153845, "end": 1315.7553221153846, "probability": 1.0 }, { "word": "的", "start": 1315.7553221153846, "end": 1315.8753221153845, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 610, "review_annotation": "", "row_id": 610, "bounded_row_id": 610, "parent_row_ids": [ 610 ], "parent_row_id": 610, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "强化学习数据的", "selected_text": "强化学习数据的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 610 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1315.8753221153845, "end": 1317.6753221153845, "text": "自动化合成与环境构建", "raw_asr_text": "自动化合成与环境构建", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "自", "start": 1315.8753221153845, "end": 1316.0553221153846, "probability": 0.99658203125 }, { "word": "动", "start": 1316.0553221153846, "end": 1316.1753221153845, "probability": 0.99951171875 }, { "word": "化", "start": 1316.1753221153845, "end": 1316.3153221153846, "probability": 0.99951171875 }, { "word": "合", "start": 1316.3153221153846, "end": 1316.5353221153846, "probability": 0.99072265625 }, { "word": "成", "start": 1316.5353221153846, "end": 1316.7353221153846, "probability": 0.9580078125 }, { "word": "与", "start": 1316.7353221153846, "end": 1316.9753221153846, "probability": 0.9990234375 }, { "word": "环", "start": 1316.9753221153846, "end": 1317.1953221153844, "probability": 1.0 }, { "word": "境", "start": 1317.1953221153844, "end": 1317.3153221153846, "probability": 1.0 }, { "word": "构", "start": 1317.3153221153846, "end": 1317.4953221153846, "probability": 0.9990234375 }, { "word": "建", "start": 1317.4953221153846, "end": 1317.6753221153845, "probability": 0.99658203125 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 611, "review_annotation": "", "row_id": 611, "bounded_row_id": 611, "parent_row_ids": [ 611 ], "parent_row_id": 611, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "自动化合成与环境构建", "selected_text": "自动化合成与环境构建", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 611 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1317.6753221153845, "end": 1318.4353221153845, "text": "说到底", "raw_asr_text": "说到底", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "说", "start": 1317.6753221153845, "end": 1318.1553221153845, "probability": 0.99365234375 }, { "word": "到底", "start": 1318.1553221153845, "end": 1318.4353221153845, "probability": 0.99658203125 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 612, "review_annotation": "", "row_id": 612, "bounded_row_id": 612, "parent_row_ids": [ 612 ], "parent_row_id": 612, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "说到底", "selected_text": "说到底", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 612 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1318.4353221153845, "end": 1319.8153221153846, "text": "我认为就是垃圾进", "raw_asr_text": "我认为就是垃圾进", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "我", "start": 1318.4353221153845, "end": 1318.6553221153845, "probability": 0.96484375 }, { "word": "认", "start": 1318.6553221153845, "end": 1318.8353221153845, "probability": 1.0 }, { "word": "为", "start": 1318.8353221153845, "end": 1318.9553221153844, "probability": 1.0 }, { "word": "就是", "start": 1318.9553221153844, "end": 1319.2353221153846, "probability": 0.99267578125 }, { "word": "垃", "start": 1319.2353221153846, "end": 1319.5353221153846, "probability": 0.9033203125 }, { "word": "圾", "start": 1319.5353221153846, "end": 1319.7153221153847, "probability": 0.9951171875 }, { "word": "进", "start": 1319.7153221153847, "end": 1319.8153221153846, "probability": 0.740234375 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 613, "review_annotation": "", "row_id": 613, "bounded_row_id": 613, "parent_row_ids": [ 613 ], "parent_row_id": 613, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "我认为就是垃圾进", "selected_text": "我认为就是垃圾进", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 613 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1319.8153221153846, "end": 1320.6553221153845, "text": "垃圾出", "raw_asr_text": "垃圾出", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "垃", "start": 1319.8153221153846, "end": 1320.2753221153846, "probability": 0.9150390625 }, { "word": "圾", "start": 1320.2753221153846, "end": 1320.4353221153845, "probability": 1.0 }, { "word": "出", "start": 1320.4353221153845, "end": 1320.6553221153845, "probability": 0.99462890625 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 614, "review_annotation": "", "row_id": 614, "bounded_row_id": 614, "parent_row_ids": [ 614 ], "parent_row_id": 614, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "垃圾出", "selected_text": "垃圾出", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 614 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1320.6553221153845, "end": 1321.6753221153845, "text": "但这也意味着", "raw_asr_text": "但这也意味着", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "但", "start": 1320.6553221153845, "end": 1321.0153221153846, "probability": 0.994140625 }, { "word": "这", "start": 1321.0153221153846, "end": 1321.2153221153847, "probability": 0.99072265625 }, { "word": "也", "start": 1321.2153221153847, "end": 1321.3153221153846, "probability": 0.99853515625 }, { "word": "意", "start": 1321.3153221153846, "end": 1321.4353221153845, "probability": 0.99951171875 }, { "word": "味", "start": 1321.4353221153845, "end": 1321.5153221153846, "probability": 0.99951171875 }, { "word": "着", "start": 1321.5153221153846, "end": 1321.6753221153845, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 615, "review_annotation": "", "row_id": 615, "bounded_row_id": 615, "parent_row_ids": [ 615 ], "parent_row_id": 615, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "但这也意味着", "selected_text": "但这也意味着", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 615 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1321.6753221153845, "end": 1322.8153221153846, "text": "当前的后训练范式", "raw_asr_text": "当前的后训练范式", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "当", "start": 1321.6753221153845, "end": 1321.8953221153845, "probability": 0.998046875 }, { "word": "前", "start": 1321.8953221153845, "end": 1322.0353221153846, "probability": 0.9990234375 }, { "word": "的", "start": 1322.0353221153846, "end": 1322.1553221153845, "probability": 0.9970703125 }, { "word": "后", "start": 1322.1553221153845, "end": 1322.2553221153846, "probability": 0.9970703125 }, { "word": "训", "start": 1322.2553221153846, "end": 1322.3753221153845, "probability": 0.9921875 }, { "word": "练", "start": 1322.3753221153845, "end": 1322.4753221153846, "probability": 0.99951171875 }, { "word": "范", "start": 1322.4753221153846, "end": 1322.6353221153845, "probability": 0.96337890625 }, { "word": "式", "start": 1322.6353221153845, "end": 1322.8153221153846, "probability": 0.81689453125 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 616, "review_annotation": "", "row_id": 616, "bounded_row_id": 616, "parent_row_ids": [ 616 ], "parent_row_id": 616, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "当前的后训练范式", "selected_text": "当前的后训练范式", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 616 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1322.8153221153846, "end": 1323.5953221153845, "text": "在一段时间内", "raw_asr_text": "在一段时间内", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 47, "words": [ { "word": "在", "start": 1322.8153221153846, "end": 1322.9953221153846, "probability": 0.99560546875 }, { "word": "一", "start": 1322.9953221153846, "end": 1323.1353221153845, "probability": 0.9990234375 }, { "word": "段", "start": 1323.1353221153845, "end": 1323.2153221153847, "probability": 1.0 }, { "word": "时间", "start": 1323.2153221153847, "end": 1323.4553221153844, "probability": 0.99853515625 }, { "word": "内", "start": 1323.4553221153844, "end": 1323.5953221153845, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.07655120849609375, "compression_ratio": 1.2877492877492878, "no_speech_prob": 5.849046941230895e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "在线策略蒸流他们现在将其称为大规模同策略蒸流V4.1会向不同领域的40多个各有专攻的教师模型学习这也说明DeepSeek目前认为投入更多时间和资金优化数据比继续研发新的后训练算法更有价值尤其是本篇论文中他们十分重视强化学习数据的自动化合成与环境构建说到底我认为就是垃圾进垃圾出但这也意味着当前的后训练范式在一段时间内可能会少很多有意思的心情" } ], "id": 617, "review_annotation": "", "row_id": 617, "bounded_row_id": 617, "parent_row_ids": [ 617 ], "parent_row_id": 617, "packet_id": 47, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1295.8353221153845, "packet_end": 1325.0233076923078, "raw_text": "在一段时间内", "selected_text": "在一段时间内", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 617 ], "packet_id": 47, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1324.2233076923078, "end": 1325.2233076923078, "text": "可能会少很多有意思的新进展", "raw_asr_text": "可能会少很多有意思的新进展", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 48, "words": [ { "word": "可能", "start": 1323.4233076923078, "end": 1323.7833076923077, "probability": 0.66650390625 }, { "word": "会", "start": 1323.7833076923077, "end": 1323.9833076923078, "probability": 0.9892578125 }, { "word": "少", "start": 1323.9833076923078, "end": 1324.123307692308, "probability": 0.99853515625 }, { "word": "很多", "start": 1324.123307692308, "end": 1324.363307692308, "probability": 0.99462890625 }, { "word": "有", "start": 1324.363307692308, "end": 1324.5233076923078, "probability": 0.99462890625 }, { "word": "意思", "start": 1324.5233076923078, "end": 1324.6833076923078, "probability": 0.998046875 }, { "word": "的", "start": 1324.6833076923078, "end": 1324.823307692308, "probability": 0.99853515625 }, { "word": "新", "start": 1324.823307692308, "end": 1324.9433076923078, "probability": 0.986328125 }, { "word": "进", "start": 1324.9433076923078, "end": 1325.063307692308, "probability": 0.99560546875 }, { "word": "展", "start": 1325.063307692308, "end": 1325.2233076923078, "probability": 0.99658203125 } ], "asr_metrics": { "avg_logprob": -0.09032595666098933, "compression_ratio": 1.4225352112676057, "no_speech_prob": 4.1161386798993504e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "可能会少很多有意思的新进展不过客观来说MOPT也是最近才推出的所以这套范式问世的时间本来也不算长不过他们还做了一项我认为之前没有明确提到过的小调整那就是模型合并我也做过一期关于模型合并的视频在强化学习过程中他们会合并来自不同脚手架不同配置的独立训练运行的检查点整合不同优化路径的提升效果在用合并后的模型初始化下一次强化学习运行也就是说模型合并" } ], "id": 618, "review_annotation": "", "row_id": 618, "bounded_row_id": 618, "parent_row_ids": [ 618 ], "parent_row_id": 618, "packet_id": 48, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1323.4233076923078, "packet_end": 1352.6112932692308, "raw_text": "可能会少很多有意思的新进展", "selected_text": "可能会少很多有意思的新进展", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 618 ], "packet_id": 48, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1325.2233076923078, "end": 1326.2433076923078, "text": "不过客观来说", "raw_asr_text": "不过客观来说", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 48, "words": [ { "word": "不", "start": 1325.2233076923078, "end": 1325.4033076923079, "probability": 0.9765625 }, { "word": "过", "start": 1325.4033076923079, "end": 1325.5433076923077, "probability": 1.0 }, { "word": "客", "start": 1325.5433076923077, "end": 1325.6833076923078, "probability": 0.99853515625 }, { "word": "观", "start": 1325.6833076923078, "end": 1325.8833076923079, "probability": 0.98779296875 }, { "word": "来", "start": 1325.8833076923079, "end": 1326.0233076923078, "probability": 0.99658203125 }, { "word": "说", "start": 1326.0233076923078, "end": 1326.2433076923078, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.09032595666098933, "compression_ratio": 1.4225352112676057, "no_speech_prob": 4.1161386798993504e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "可能会少很多有意思的新进展不过客观来说MOPT也是最近才推出的所以这套范式问世的时间本来也不算长不过他们还做了一项我认为之前没有明确提到过的小调整那就是模型合并我也做过一期关于模型合并的视频在强化学习过程中他们会合并来自不同脚手架不同配置的独立训练运行的检查点整合不同优化路径的提升效果在用合并后的模型初始化下一次强化学习运行也就是说模型合并" } ], "id": 619, "review_annotation": "", "row_id": 619, "bounded_row_id": 619, "parent_row_ids": [ 619 ], "parent_row_id": 619, "packet_id": 48, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1323.4233076923078, "packet_end": 1352.6112932692308, "raw_text": "不过客观来说", "selected_text": "不过客观来说", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 619 ], "packet_id": 48, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1326.2433076923078, "end": 1328.123307692308, "text": "MOPT也是最近才推出的", "raw_asr_text": "MOPT也是最近才推出的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 48, "words": [ { "word": "M", "start": 1326.2433076923078, "end": 1326.4433076923078, "probability": 0.67626953125 }, { "word": "OP", "start": 1326.4433076923078, "end": 1326.5233076923078, "probability": 0.1717529296875 }, { "word": "T", "start": 1326.5233076923078, "end": 1326.6833076923078, "probability": 0.48388671875 }, { "word": "也是", "start": 1326.6833076923078, "end": 1327.0033076923078, "probability": 0.91357421875 }, { "word": "最近", "start": 1327.0033076923078, "end": 1327.4033076923079, "probability": 0.998046875 }, { "word": "才", "start": 1327.4033076923079, "end": 1327.583307692308, "probability": 0.99755859375 }, { "word": "推", "start": 1327.583307692308, "end": 1327.7833076923077, "probability": 0.99755859375 }, { "word": "出", "start": 1327.7833076923077, "end": 1327.9833076923078, "probability": 0.99853515625 }, { "word": "的", "start": 1327.9833076923078, "end": 1328.123307692308, "probability": 0.9951171875 } ], "asr_metrics": { "avg_logprob": -0.09032595666098933, "compression_ratio": 1.4225352112676057, "no_speech_prob": 4.1161386798993504e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "可能会少很多有意思的新进展不过客观来说MOPT也是最近才推出的所以这套范式问世的时间本来也不算长不过他们还做了一项我认为之前没有明确提到过的小调整那就是模型合并我也做过一期关于模型合并的视频在强化学习过程中他们会合并来自不同脚手架不同配置的独立训练运行的检查点整合不同优化路径的提升效果在用合并后的模型初始化下一次强化学习运行也就是说模型合并" } ], "id": 620, "review_annotation": "", "row_id": 620, "bounded_row_id": 620, "parent_row_ids": [ 620 ], "parent_row_id": 620, "packet_id": 48, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1323.4233076923078, "packet_end": 1352.6112932692308, "raw_text": "MOPT也是最近才推出的", "selected_text": "MOPT也是最近才推出的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 620 ], "packet_id": 48, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1328.123307692308, "end": 1329.9433076923078, "text": "所以这套范式问世的时间", "raw_asr_text": "所以这套范式问世的时间", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 48, "words": [ { "word": "所以", "start": 1328.123307692308, "end": 1328.4433076923078, "probability": 0.99365234375 }, { "word": "这", "start": 1328.4433076923078, "end": 1328.623307692308, "probability": 0.9267578125 }, { "word": "套", "start": 1328.623307692308, "end": 1328.7833076923077, "probability": 0.994140625 }, { "word": "范", "start": 1328.7833076923077, "end": 1329.0233076923078, "probability": 0.72705078125 }, { "word": "式", "start": 1329.0233076923078, "end": 1329.2233076923078, "probability": 0.9482421875 }, { "word": "问", "start": 1329.2233076923078, "end": 1329.4033076923079, "probability": 0.98291015625 }, { "word": "世", "start": 1329.4033076923079, "end": 1329.563307692308, "probability": 0.83935546875 }, { "word": "的", "start": 1329.563307692308, "end": 1329.6833076923078, "probability": 0.99658203125 }, { "word": "时间", "start": 1329.6833076923078, "end": 1329.9433076923078, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.09032595666098933, "compression_ratio": 1.4225352112676057, "no_speech_prob": 4.1161386798993504e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "可能会少很多有意思的新进展不过客观来说MOPT也是最近才推出的所以这套范式问世的时间本来也不算长不过他们还做了一项我认为之前没有明确提到过的小调整那就是模型合并我也做过一期关于模型合并的视频在强化学习过程中他们会合并来自不同脚手架不同配置的独立训练运行的检查点整合不同优化路径的提升效果在用合并后的模型初始化下一次强化学习运行也就是说模型合并" } ], "id": 621, "review_annotation": "", "row_id": 621, "bounded_row_id": 621, "parent_row_ids": [ 621 ], "parent_row_id": 621, "packet_id": 48, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1323.4233076923078, "packet_end": 1352.6112932692308, "raw_text": "所以这套范式问世的时间", "selected_text": "所以这套范式问世的时间", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 621 ], "packet_id": 48, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1329.9433076923078, "end": 1331.123307692308, "text": "本来也不算长", "raw_asr_text": "本来也不算长", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 48, "words": [ { "word": "本", "start": 1329.9433076923078, "end": 1330.2433076923078, "probability": 0.98876953125 }, { "word": "来", "start": 1330.2433076923078, "end": 1330.4233076923078, "probability": 0.99951171875 }, { "word": "也", "start": 1330.4233076923078, "end": 1330.583307692308, "probability": 0.99853515625 }, { "word": "不", "start": 1330.583307692308, "end": 1330.7433076923078, "probability": 0.998046875 }, { "word": "算", "start": 1330.7433076923078, "end": 1330.9233076923078, "probability": 1.0 }, { "word": "长", "start": 1330.9233076923078, "end": 1331.123307692308, "probability": 0.9892578125 } ], "asr_metrics": { "avg_logprob": -0.09032595666098933, "compression_ratio": 1.4225352112676057, "no_speech_prob": 4.1161386798993504e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "可能会少很多有意思的新进展不过客观来说MOPT也是最近才推出的所以这套范式问世的时间本来也不算长不过他们还做了一项我认为之前没有明确提到过的小调整那就是模型合并我也做过一期关于模型合并的视频在强化学习过程中他们会合并来自不同脚手架不同配置的独立训练运行的检查点整合不同优化路径的提升效果在用合并后的模型初始化下一次强化学习运行也就是说模型合并" } ], "id": 622, "review_annotation": "", "row_id": 622, "bounded_row_id": 622, "parent_row_ids": [ 622 ], "parent_row_id": 622, "packet_id": 48, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1323.4233076923078, "packet_end": 1352.6112932692308, "raw_text": "本来也不算长", "selected_text": "本来也不算长", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 622 ], "packet_id": 48, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1331.123307692308, "end": 1332.4233076923078, "text": "不过他们还做了一项", "raw_asr_text": "不过他们还做了一项", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 48, "words": [ { "word": "不", "start": 1331.123307692308, "end": 1331.4233076923078, "probability": 0.98974609375 }, { "word": "过", "start": 1331.4233076923078, "end": 1331.5233076923078, "probability": 1.0 }, { "word": "他们", "start": 1331.5233076923078, "end": 1331.7033076923078, "probability": 0.978515625 }, { "word": "还", "start": 1331.7033076923078, "end": 1331.863307692308, "probability": 0.9990234375 }, { "word": "做", "start": 1331.863307692308, "end": 1332.0433076923077, "probability": 0.99853515625 }, { "word": "了", "start": 1332.0433076923077, "end": 1332.1633076923079, "probability": 0.99951171875 }, { "word": "一", "start": 1332.1633076923079, "end": 1332.2833076923077, "probability": 0.9853515625 }, { "word": "项", "start": 1332.2833076923077, "end": 1332.4233076923078, "probability": 0.998046875 } ], "asr_metrics": { "avg_logprob": -0.09032595666098933, "compression_ratio": 1.4225352112676057, "no_speech_prob": 4.1161386798993504e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "可能会少很多有意思的新进展不过客观来说MOPT也是最近才推出的所以这套范式问世的时间本来也不算长不过他们还做了一项我认为之前没有明确提到过的小调整那就是模型合并我也做过一期关于模型合并的视频在强化学习过程中他们会合并来自不同脚手架不同配置的独立训练运行的检查点整合不同优化路径的提升效果在用合并后的模型初始化下一次强化学习运行也就是说模型合并" } ], "id": 623, "review_annotation": "", "row_id": 623, "bounded_row_id": 623, "parent_row_ids": [ 623 ], "parent_row_id": 623, "packet_id": 48, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1323.4233076923078, "packet_end": 1352.6112932692308, "raw_text": "不过他们还做了一项", "selected_text": "不过他们还做了一项", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 623 ], "packet_id": 48, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1332.4233076923078, "end": 1334.8833076923079, "text": "我认为之前没有明确提到过的小调整", "raw_asr_text": "我认为之前没有明确提到过的小调整", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 48, "words": [ { "word": "我", "start": 1332.4233076923078, "end": 1332.583307692308, "probability": 0.99853515625 }, { "word": "认", "start": 1332.583307692308, "end": 1332.7433076923078, "probability": 0.99951171875 }, { "word": "为", "start": 1332.7433076923078, "end": 1332.863307692308, "probability": 0.9990234375 }, { "word": "之前", "start": 1332.863307692308, "end": 1333.1833076923078, "probability": 0.994140625 }, { "word": "没有", "start": 1333.1833076923078, "end": 1333.4433076923078, "probability": 0.99755859375 }, { "word": "明", "start": 1333.4433076923078, "end": 1333.623307692308, "probability": 0.99951171875 }, { "word": "确", "start": 1333.623307692308, "end": 1333.863307692308, "probability": 1.0 }, { "word": "提", "start": 1333.863307692308, "end": 1334.0033076923078, "probability": 0.99951171875 }, { "word": "到", "start": 1334.0033076923078, "end": 1334.1633076923079, "probability": 0.99853515625 }, { "word": "过", "start": 1334.1633076923079, "end": 1334.2833076923077, "probability": 0.9873046875 }, { "word": "的", "start": 1334.2833076923077, "end": 1334.363307692308, "probability": 0.99853515625 }, { "word": "小", "start": 1334.363307692308, "end": 1334.5033076923078, "probability": 0.99853515625 }, { "word": "调", "start": 1334.5033076923078, "end": 1334.7233076923078, "probability": 0.9990234375 }, { "word": "整", "start": 1334.7233076923078, "end": 1334.8833076923079, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09032595666098933, "compression_ratio": 1.4225352112676057, "no_speech_prob": 4.1161386798993504e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "可能会少很多有意思的新进展不过客观来说MOPT也是最近才推出的所以这套范式问世的时间本来也不算长不过他们还做了一项我认为之前没有明确提到过的小调整那就是模型合并我也做过一期关于模型合并的视频在强化学习过程中他们会合并来自不同脚手架不同配置的独立训练运行的检查点整合不同优化路径的提升效果在用合并后的模型初始化下一次强化学习运行也就是说模型合并" } ], "id": 624, "review_annotation": "", "row_id": 624, "bounded_row_id": 624, "parent_row_ids": [ 624 ], "parent_row_id": 624, "packet_id": 48, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1323.4233076923078, "packet_end": 1352.6112932692308, "raw_text": "我认为之前没有明确提到过的小调整", "selected_text": "我认为之前没有明确提到过的小调整", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 624 ], "packet_id": 48, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1334.8833076923079, "end": 1336.4433076923078, "text": "那就是模型合并", "raw_asr_text": "那就是模型合并", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 48, "words": [ { "word": "那", "start": 1334.8833076923079, "end": 1335.2433076923078, "probability": 0.98486328125 }, { "word": "就是", "start": 1335.2433076923078, "end": 1335.5233076923078, "probability": 0.9970703125 }, { "word": "模", "start": 1335.5233076923078, "end": 1335.823307692308, "probability": 0.95947265625 }, { "word": "型", "start": 1335.823307692308, "end": 1336.0433076923077, "probability": 0.99951171875 }, { "word": "合", "start": 1336.0433076923077, "end": 1336.2433076923078, "probability": 0.994140625 }, { "word": "并", "start": 1336.2433076923078, "end": 1336.4433076923078, "probability": 0.9931640625 } ], "asr_metrics": { "avg_logprob": -0.09032595666098933, "compression_ratio": 1.4225352112676057, "no_speech_prob": 4.1161386798993504e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "可能会少很多有意思的新进展不过客观来说MOPT也是最近才推出的所以这套范式问世的时间本来也不算长不过他们还做了一项我认为之前没有明确提到过的小调整那就是模型合并我也做过一期关于模型合并的视频在强化学习过程中他们会合并来自不同脚手架不同配置的独立训练运行的检查点整合不同优化路径的提升效果在用合并后的模型初始化下一次强化学习运行也就是说模型合并" } ], "id": 625, "review_annotation": "", "row_id": 625, "bounded_row_id": 625, "parent_row_ids": [ 625 ], "parent_row_id": 625, "packet_id": 48, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1323.4233076923078, "packet_end": 1352.6112932692308, "raw_text": "那就是模型合并", "selected_text": "那就是模型合并", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 625 ], "packet_id": 48, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1336.4433076923078, "end": 1338.623307692308, "text": "我也做过一期关于模型合并的视频", "raw_asr_text": "我也做过一期关于模型合并的视频", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 48, "words": [ { "word": "我也", "start": 1336.4433076923078, "end": 1336.9833076923078, "probability": 0.9775390625 }, { "word": "做", "start": 1336.9833076923078, "end": 1337.123307692308, "probability": 0.99853515625 }, { "word": "过", "start": 1337.123307692308, "end": 1337.2433076923078, "probability": 0.9990234375 }, { "word": "一", "start": 1337.2433076923078, "end": 1337.363307692308, "probability": 0.9755859375 }, { "word": "期", "start": 1337.363307692308, "end": 1337.4633076923078, "probability": 0.99951171875 }, { "word": "关", "start": 1337.4633076923078, "end": 1337.603307692308, "probability": 0.99560546875 }, { "word": "于", "start": 1337.603307692308, "end": 1337.7233076923078, "probability": 0.99853515625 }, { "word": "模", "start": 1337.7233076923078, "end": 1337.843307692308, "probability": 0.99365234375 }, { "word": "型", "start": 1337.843307692308, "end": 1338.0033076923078, "probability": 0.99951171875 }, { "word": "合", "start": 1338.0033076923078, "end": 1338.1433076923079, "probability": 0.99560546875 }, { "word": "并", "start": 1338.1433076923079, "end": 1338.2833076923077, "probability": 0.998046875 }, { "word": "的", "start": 1338.2833076923077, "end": 1338.3833076923079, "probability": 0.998046875 }, { "word": "视", "start": 1338.3833076923079, "end": 1338.4833076923078, "probability": 0.998046875 }, { "word": "频", "start": 1338.4833076923078, "end": 1338.623307692308, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09032595666098933, "compression_ratio": 1.4225352112676057, "no_speech_prob": 4.1161386798993504e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "可能会少很多有意思的新进展不过客观来说MOPT也是最近才推出的所以这套范式问世的时间本来也不算长不过他们还做了一项我认为之前没有明确提到过的小调整那就是模型合并我也做过一期关于模型合并的视频在强化学习过程中他们会合并来自不同脚手架不同配置的独立训练运行的检查点整合不同优化路径的提升效果在用合并后的模型初始化下一次强化学习运行也就是说模型合并" } ], "id": 626, "review_annotation": "", "row_id": 626, "bounded_row_id": 626, "parent_row_ids": [ 626 ], "parent_row_id": 626, "packet_id": 48, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1323.4233076923078, "packet_end": 1352.6112932692308, "raw_text": "我也做过一期关于模型合并的视频", "selected_text": "我也做过一期关于模型合并的视频", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 626 ], "packet_id": 48, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1338.623307692308, "end": 1340.303307692308, "text": "在强化学习过程中", "raw_asr_text": "在强化学习过程中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 48, "words": [ { "word": "在", "start": 1338.623307692308, "end": 1339.0233076923078, "probability": 0.99365234375 }, { "word": "强", "start": 1339.0233076923078, "end": 1339.323307692308, "probability": 0.9990234375 }, { "word": "化", "start": 1339.323307692308, "end": 1339.5433076923077, "probability": 0.99951171875 }, { "word": "学", "start": 1339.5433076923077, "end": 1339.7033076923078, "probability": 1.0 }, { "word": "习", "start": 1339.7033076923078, "end": 1339.843307692308, "probability": 1.0 }, { "word": "过", "start": 1339.843307692308, "end": 1339.9633076923078, "probability": 0.99658203125 }, { "word": "程", "start": 1339.9633076923078, "end": 1340.123307692308, "probability": 1.0 }, { "word": "中", "start": 1340.123307692308, "end": 1340.303307692308, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.09032595666098933, "compression_ratio": 1.4225352112676057, "no_speech_prob": 4.1161386798993504e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "可能会少很多有意思的新进展不过客观来说MOPT也是最近才推出的所以这套范式问世的时间本来也不算长不过他们还做了一项我认为之前没有明确提到过的小调整那就是模型合并我也做过一期关于模型合并的视频在强化学习过程中他们会合并来自不同脚手架不同配置的独立训练运行的检查点整合不同优化路径的提升效果在用合并后的模型初始化下一次强化学习运行也就是说模型合并" } ], "id": 627, "review_annotation": "", "row_id": 627, "bounded_row_id": 627, "parent_row_ids": [ 627 ], "parent_row_id": 627, "packet_id": 48, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1323.4233076923078, "packet_end": 1352.6112932692308, "raw_text": "在强化学习过程中", "selected_text": "在强化学习过程中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 627 ], "packet_id": 48, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1340.303307692308, "end": 1342.6633076923079, "text": "他们会合并来自不同脚手架", "raw_asr_text": "他们会合并来自不同脚手架", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 48, "words": [ { "word": "他们", "start": 1340.303307692308, "end": 1340.603307692308, "probability": 0.84912109375 }, { "word": "会", "start": 1340.603307692308, "end": 1340.803307692308, "probability": 0.99951171875 }, { "word": "合", "start": 1340.803307692308, "end": 1341.083307692308, "probability": 0.9970703125 }, { "word": "并", "start": 1341.083307692308, "end": 1341.2833076923077, "probability": 0.99951171875 }, { "word": "来", "start": 1341.2833076923077, "end": 1341.4633076923078, "probability": 0.99853515625 }, { "word": "自", "start": 1341.4633076923078, "end": 1341.6433076923079, "probability": 0.99951171875 }, { "word": "不同", "start": 1341.6433076923079, "end": 1342.0233076923078, "probability": 0.9970703125 }, { "word": "脚", "start": 1342.0233076923078, "end": 1342.2633076923078, "probability": 0.9833984375 }, { "word": "手", "start": 1342.2633076923078, "end": 1342.4433076923078, "probability": 0.98974609375 }, { "word": "架", "start": 1342.4433076923078, "end": 1342.6633076923079, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.09032595666098933, "compression_ratio": 1.4225352112676057, "no_speech_prob": 4.1161386798993504e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "可能会少很多有意思的新进展不过客观来说MOPT也是最近才推出的所以这套范式问世的时间本来也不算长不过他们还做了一项我认为之前没有明确提到过的小调整那就是模型合并我也做过一期关于模型合并的视频在强化学习过程中他们会合并来自不同脚手架不同配置的独立训练运行的检查点整合不同优化路径的提升效果在用合并后的模型初始化下一次强化学习运行也就是说模型合并" } ], "id": 628, "review_annotation": "", "row_id": 628, "bounded_row_id": 628, "parent_row_ids": [ 628 ], "parent_row_id": 628, "packet_id": 48, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1323.4233076923078, "packet_end": 1352.6112932692308, "raw_text": "他们会合并来自不同脚手架", "selected_text": "他们会合并来自不同脚手架", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 628 ], "packet_id": 48, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1342.6633076923079, "end": 1345.2633076923078, "text": "不同配置的独立训练运行的检查点", "raw_asr_text": "不同配置的独立训练运行的检查点", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 48, "words": [ { "word": "不同", "start": 1342.6633076923079, "end": 1342.9633076923078, "probability": 0.9443359375 }, { "word": "配", "start": 1342.9633076923078, "end": 1343.1833076923078, "probability": 0.99755859375 }, { "word": "置", "start": 1343.1833076923078, "end": 1343.363307692308, "probability": 0.99853515625 }, { "word": "的", "start": 1343.363307692308, "end": 1343.5033076923078, "probability": 0.99755859375 }, { "word": "独", "start": 1343.5033076923078, "end": 1343.7833076923077, "probability": 0.998046875 }, { "word": "立", "start": 1343.7833076923077, "end": 1343.9633076923078, "probability": 0.9990234375 }, { "word": "训", "start": 1343.9633076923078, "end": 1344.2033076923078, "probability": 0.9990234375 }, { "word": "练", "start": 1344.2033076923078, "end": 1344.363307692308, "probability": 0.99951171875 }, { "word": "运", "start": 1344.363307692308, "end": 1344.5033076923078, "probability": 1.0 }, { "word": "行", "start": 1344.5033076923078, "end": 1344.623307692308, "probability": 0.99609375 }, { "word": "的", "start": 1344.623307692308, "end": 1344.7433076923078, "probability": 0.998046875 }, { "word": "检", "start": 1344.7433076923078, "end": 1344.9233076923078, "probability": 0.9990234375 }, { "word": "查", "start": 1344.9233076923078, "end": 1345.083307692308, "probability": 0.98876953125 }, { "word": "点", "start": 1345.083307692308, "end": 1345.2633076923078, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.09032595666098933, "compression_ratio": 1.4225352112676057, "no_speech_prob": 4.1161386798993504e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "可能会少很多有意思的新进展不过客观来说MOPT也是最近才推出的所以这套范式问世的时间本来也不算长不过他们还做了一项我认为之前没有明确提到过的小调整那就是模型合并我也做过一期关于模型合并的视频在强化学习过程中他们会合并来自不同脚手架不同配置的独立训练运行的检查点整合不同优化路径的提升效果在用合并后的模型初始化下一次强化学习运行也就是说模型合并" } ], "id": 629, "review_annotation": "", "row_id": 629, "bounded_row_id": 629, "parent_row_ids": [ 629 ], "parent_row_id": 629, "packet_id": 48, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1323.4233076923078, "packet_end": 1352.6112932692308, "raw_text": "不同配置的独立训练运行的检查点", "selected_text": "不同配置的独立训练运行的检查点", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 629 ], "packet_id": 48, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1345.2633076923078, "end": 1347.6833076923078, "text": "整合不同优化路径的提升效果", "raw_asr_text": "整合不同优化路径的提升效果", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 48, "words": [ { "word": "整", "start": 1345.2633076923078, "end": 1345.6433076923079, "probability": 0.998046875 }, { "word": "合", "start": 1345.6433076923079, "end": 1345.9033076923079, "probability": 0.99853515625 }, { "word": "不同", "start": 1345.9033076923079, "end": 1346.2033076923078, "probability": 0.99853515625 }, { "word": "优", "start": 1346.2033076923078, "end": 1346.4033076923079, "probability": 0.9990234375 }, { "word": "化", "start": 1346.4033076923079, "end": 1346.563307692308, "probability": 0.99951171875 }, { "word": "路", "start": 1346.563307692308, "end": 1346.7033076923078, "probability": 0.99755859375 }, { "word": "径", "start": 1346.7033076923078, "end": 1346.7833076923077, "probability": 1.0 }, { "word": "的", "start": 1346.7833076923077, "end": 1346.9033076923079, "probability": 0.99853515625 }, { "word": "提", "start": 1346.9033076923079, "end": 1347.063307692308, "probability": 1.0 }, { "word": "升", "start": 1347.063307692308, "end": 1347.2433076923078, "probability": 0.99853515625 }, { "word": "效", "start": 1347.2433076923078, "end": 1347.4233076923078, "probability": 0.998046875 }, { "word": "果", "start": 1347.4233076923078, "end": 1347.6833076923078, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.09032595666098933, "compression_ratio": 1.4225352112676057, "no_speech_prob": 4.1161386798993504e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "可能会少很多有意思的新进展不过客观来说MOPT也是最近才推出的所以这套范式问世的时间本来也不算长不过他们还做了一项我认为之前没有明确提到过的小调整那就是模型合并我也做过一期关于模型合并的视频在强化学习过程中他们会合并来自不同脚手架不同配置的独立训练运行的检查点整合不同优化路径的提升效果在用合并后的模型初始化下一次强化学习运行也就是说模型合并" } ], "id": 630, "review_annotation": "", "row_id": 630, "bounded_row_id": 630, "parent_row_ids": [ 630 ], "parent_row_id": 630, "packet_id": 48, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1323.4233076923078, "packet_end": 1352.6112932692308, "raw_text": "整合不同优化路径的提升效果", "selected_text": "整合不同优化路径的提升效果", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 630 ], "packet_id": 48, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1347.6833076923078, "end": 1349.6433076923079, "text": "在用合并后的模型初始化", "raw_asr_text": "在用合并后的模型初始化", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 48, "words": [ { "word": "在", "start": 1347.6833076923078, "end": 1348.2433076923078, "probability": 0.51025390625 }, { "word": "用", "start": 1348.2433076923078, "end": 1348.363307692308, "probability": 0.99853515625 }, { "word": "合", "start": 1348.363307692308, "end": 1348.5233076923078, "probability": 0.98095703125 }, { "word": "并", "start": 1348.5233076923078, "end": 1348.6433076923079, "probability": 0.99609375 }, { "word": "后", "start": 1348.6433076923079, "end": 1348.7433076923078, "probability": 0.99560546875 }, { "word": "的", "start": 1348.7433076923078, "end": 1348.823307692308, "probability": 0.99951171875 }, { "word": "模", "start": 1348.823307692308, "end": 1348.9633076923078, "probability": 0.998046875 }, { "word": "型", "start": 1348.9633076923078, "end": 1349.1633076923079, "probability": 0.99951171875 }, { "word": "初", "start": 1349.1633076923079, "end": 1349.363307692308, "probability": 0.990234375 }, { "word": "始", "start": 1349.363307692308, "end": 1349.5033076923078, "probability": 0.9853515625 }, { "word": "化", "start": 1349.5033076923078, "end": 1349.6433076923079, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.09032595666098933, "compression_ratio": 1.4225352112676057, "no_speech_prob": 4.1161386798993504e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "可能会少很多有意思的新进展不过客观来说MOPT也是最近才推出的所以这套范式问世的时间本来也不算长不过他们还做了一项我认为之前没有明确提到过的小调整那就是模型合并我也做过一期关于模型合并的视频在强化学习过程中他们会合并来自不同脚手架不同配置的独立训练运行的检查点整合不同优化路径的提升效果在用合并后的模型初始化下一次强化学习运行也就是说模型合并" } ], "id": 631, "review_annotation": "", "row_id": 631, "bounded_row_id": 631, "parent_row_ids": [ 631 ], "parent_row_id": 631, "packet_id": 48, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1323.4233076923078, "packet_end": 1352.6112932692308, "raw_text": "在用合并后的模型初始化", "selected_text": "在用合并后的模型初始化", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 631 ], "packet_id": 48, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1349.6433076923079, "end": 1351.063307692308, "text": "下一次强化学习运行", "raw_asr_text": "下一次强化学习运行", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 48, "words": [ { "word": "下", "start": 1349.6433076923079, "end": 1349.823307692308, "probability": 0.994140625 }, { "word": "一次", "start": 1349.823307692308, "end": 1350.0433076923077, "probability": 0.998046875 }, { "word": "强", "start": 1350.0433076923077, "end": 1350.2633076923078, "probability": 0.9990234375 }, { "word": "化", "start": 1350.2633076923078, "end": 1350.4433076923078, "probability": 0.99951171875 }, { "word": "学", "start": 1350.4433076923078, "end": 1350.603307692308, "probability": 0.9990234375 }, { "word": "习", "start": 1350.603307692308, "end": 1350.7233076923078, "probability": 1.0 }, { "word": "运", "start": 1350.7233076923078, "end": 1350.8833076923079, "probability": 1.0 }, { "word": "行", "start": 1350.8833076923079, "end": 1351.063307692308, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.09032595666098933, "compression_ratio": 1.4225352112676057, "no_speech_prob": 4.1161386798993504e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "可能会少很多有意思的新进展不过客观来说MOPT也是最近才推出的所以这套范式问世的时间本来也不算长不过他们还做了一项我认为之前没有明确提到过的小调整那就是模型合并我也做过一期关于模型合并的视频在强化学习过程中他们会合并来自不同脚手架不同配置的独立训练运行的检查点整合不同优化路径的提升效果在用合并后的模型初始化下一次强化学习运行也就是说模型合并" } ], "id": 632, "review_annotation": "", "row_id": 632, "bounded_row_id": 632, "parent_row_ids": [ 632 ], "parent_row_id": 632, "packet_id": 48, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1323.4233076923078, "packet_end": 1352.6112932692308, "raw_text": "下一次强化学习运行", "selected_text": "下一次强化学习运行", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 632 ], "packet_id": 48, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1351.8112932692309, "end": 1354.671293269231, "text": "也就是说模型合并是用来持续汇总各次成功", "raw_asr_text": "也就是说模型合并是用来持续汇总各次成功", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 49, "words": [ { "word": "也", "start": 1351.011293269231, "end": 1351.3312932692309, "probability": 0.9541015625 }, { "word": "就是", "start": 1351.3312932692309, "end": 1351.511293269231, "probability": 0.9931640625 }, { "word": "说", "start": 1351.511293269231, "end": 1351.691293269231, "probability": 0.99072265625 }, { "word": "模", "start": 1351.691293269231, "end": 1351.931293269231, "probability": 0.88671875 }, { "word": "型", "start": 1351.931293269231, "end": 1352.1312932692308, "probability": 0.9990234375 }, { "word": "合", "start": 1352.1312932692308, "end": 1352.411293269231, "probability": 0.99560546875 }, { "word": "并", "start": 1352.411293269231, "end": 1352.5912932692308, "probability": 0.98828125 }, { "word": "是", "start": 1352.5912932692308, "end": 1352.791293269231, "probability": 0.96484375 }, { "word": "用", "start": 1352.791293269231, "end": 1352.911293269231, "probability": 0.998046875 }, { "word": "来", "start": 1352.911293269231, "end": 1353.0512932692309, "probability": 0.99609375 }, { "word": "持", "start": 1353.0512932692309, "end": 1353.3112932692309, "probability": 0.99609375 }, { "word": "续", "start": 1353.3112932692309, "end": 1353.511293269231, "probability": 1.0 }, { "word": "汇", "start": 1353.511293269231, "end": 1353.6312932692308, "probability": 0.9990234375 }, { "word": "总", "start": 1353.6312932692308, "end": 1353.771293269231, "probability": 0.9970703125 }, { "word": "各", "start": 1353.771293269231, "end": 1353.971293269231, "probability": 0.9765625 }, { "word": "次", "start": 1353.971293269231, "end": 1354.171293269231, "probability": 0.98876953125 }, { "word": "成", "start": 1354.171293269231, "end": 1354.3712932692308, "probability": 0.9970703125 }, { "word": "功", "start": 1354.3712932692308, "end": 1354.671293269231, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.0805326558775821, "compression_ratio": 1.2604790419161678, "no_speech_prob": 2.169375790117556e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "也就是说模型合并是用来持续汇总各次成功强化学习运行带来的提升等到流程最后他们会在不同领域和训练阶段开展大规模同策略蒸馏将所有专家教师模型的能力整合到同一个模型中另一方面推理力度也有一项升级它不再像V4那样只设置Think High和Think Max模式而是通过强化学习让推理力度成为一个从1到100的数值力度越低Token" } ], "id": 633, "review_annotation": "", "row_id": 633, "bounded_row_id": 633, "parent_row_ids": [ 633 ], "parent_row_id": 633, "packet_id": 49, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1351.011293269231, "packet_end": 1380.1992788461541, "raw_text": "也就是说模型合并是用来持续汇总各次成功", "selected_text": "也就是说模型合并是用来持续汇总各次成功", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 633 ], "packet_id": 49, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1354.671293269231, "end": 1357.671293269231, "text": "强化学习运行带来的提升等到流程最后", "raw_asr_text": "强化学习运行带来的提升等到流程最后", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 49, "words": [ { "word": "强", "start": 1354.671293269231, "end": 1355.0712932692309, "probability": 0.9287109375 }, { "word": "化", "start": 1355.0712932692309, "end": 1355.271293269231, "probability": 0.99951171875 }, { "word": "学", "start": 1355.271293269231, "end": 1355.451293269231, "probability": 0.998046875 }, { "word": "习", "start": 1355.451293269231, "end": 1355.6112932692308, "probability": 0.99951171875 }, { "word": "运", "start": 1355.6112932692308, "end": 1355.751293269231, "probability": 0.99951171875 }, { "word": "行", "start": 1355.751293269231, "end": 1355.931293269231, "probability": 0.99853515625 }, { "word": "带", "start": 1355.931293269231, "end": 1356.1312932692308, "probability": 0.990234375 }, { "word": "来", "start": 1356.1312932692308, "end": 1356.251293269231, "probability": 0.99951171875 }, { "word": "的", "start": 1356.251293269231, "end": 1356.3512932692308, "probability": 0.99755859375 }, { "word": "提", "start": 1356.3512932692308, "end": 1356.5512932692309, "probability": 0.99951171875 }, { "word": "升", "start": 1356.5512932692309, "end": 1356.731293269231, "probability": 0.9990234375 }, { "word": "等", "start": 1356.731293269231, "end": 1356.911293269231, "probability": 0.9892578125 }, { "word": "到", "start": 1356.911293269231, "end": 1357.0512932692309, "probability": 0.99658203125 }, { "word": "流", "start": 1357.0512932692309, "end": 1357.191293269231, "probability": 0.97900390625 }, { "word": "程", "start": 1357.191293269231, "end": 1357.3312932692309, "probability": 0.99853515625 }, { "word": "最", "start": 1357.3312932692309, "end": 1357.531293269231, "probability": 0.990234375 }, { "word": "后", "start": 1357.531293269231, "end": 1357.671293269231, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.0805326558775821, "compression_ratio": 1.2604790419161678, "no_speech_prob": 2.169375790117556e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "也就是说模型合并是用来持续汇总各次成功强化学习运行带来的提升等到流程最后他们会在不同领域和训练阶段开展大规模同策略蒸馏将所有专家教师模型的能力整合到同一个模型中另一方面推理力度也有一项升级它不再像V4那样只设置Think High和Think Max模式而是通过强化学习让推理力度成为一个从1到100的数值力度越低Token" } ], "id": 634, "review_annotation": "", "row_id": 634, "bounded_row_id": 634, "parent_row_ids": [ 634 ], "parent_row_id": 634, "packet_id": 49, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1351.011293269231, "packet_end": 1380.1992788461541, "raw_text": "强化学习运行带来的提升等到流程最后", "selected_text": "强化学习运行带来的提升等到流程最后", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 634 ], "packet_id": 49, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1357.671293269231, "end": 1361.8712932692308, "text": "他们会在不同领域和训练阶段开展大规模同策略蒸馏", "raw_asr_text": "他们会在不同领域和训练阶段开展大规模同策略蒸馏", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 49, "words": [ { "word": "他们", "start": 1357.671293269231, "end": 1358.031293269231, "probability": 0.6904296875 }, { "word": "会", "start": 1358.031293269231, "end": 1358.251293269231, "probability": 0.9970703125 }, { "word": "在", "start": 1358.251293269231, "end": 1358.451293269231, "probability": 0.9990234375 }, { "word": "不同", "start": 1358.451293269231, "end": 1358.751293269231, "probability": 0.99853515625 }, { "word": "领", "start": 1358.751293269231, "end": 1359.011293269231, "probability": 1.0 }, { "word": "域", "start": 1359.011293269231, "end": 1359.171293269231, "probability": 1.0 }, { "word": "和", "start": 1359.171293269231, "end": 1359.3312932692309, "probability": 0.9970703125 }, { "word": "训", "start": 1359.3312932692309, "end": 1359.471293269231, "probability": 0.9990234375 }, { "word": "练", "start": 1359.471293269231, "end": 1359.5912932692308, "probability": 0.99951171875 }, { "word": "阶", "start": 1359.5912932692308, "end": 1359.751293269231, "probability": 0.9990234375 }, { "word": "段", "start": 1359.751293269231, "end": 1359.931293269231, "probability": 1.0 }, { "word": "开", "start": 1359.931293269231, "end": 1360.171293269231, "probability": 0.986328125 }, { "word": "展", "start": 1360.171293269231, "end": 1360.3512932692308, "probability": 1.0 }, { "word": "大", "start": 1360.3512932692308, "end": 1360.5712932692309, "probability": 0.99853515625 }, { "word": "规", "start": 1360.5712932692309, "end": 1360.771293269231, "probability": 1.0 }, { "word": "模", "start": 1360.771293269231, "end": 1360.951293269231, "probability": 0.99951171875 }, { "word": "同", "start": 1360.951293269231, "end": 1361.151293269231, "probability": 0.9638671875 }, { "word": "策", "start": 1361.151293269231, "end": 1361.391293269231, "probability": 0.9990234375 }, { "word": "略", "start": 1361.391293269231, "end": 1361.5512932692309, "probability": 1.0 }, { "word": "蒸", "start": 1361.5512932692309, "end": 1361.691293269231, "probability": 0.837890625 }, { "word": "馏", "start": 1361.691293269231, "end": 1361.8712932692308, "probability": 0.7646484375 } ], "asr_metrics": { "avg_logprob": -0.0805326558775821, "compression_ratio": 1.2604790419161678, "no_speech_prob": 2.169375790117556e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "也就是说模型合并是用来持续汇总各次成功强化学习运行带来的提升等到流程最后他们会在不同领域和训练阶段开展大规模同策略蒸馏将所有专家教师模型的能力整合到同一个模型中另一方面推理力度也有一项升级它不再像V4那样只设置Think High和Think Max模式而是通过强化学习让推理力度成为一个从1到100的数值力度越低Token" } ], "id": 635, "review_annotation": "", "row_id": 635, "bounded_row_id": 635, "parent_row_ids": [ 635 ], "parent_row_id": 635, "packet_id": 49, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1351.011293269231, "packet_end": 1380.1992788461541, "raw_text": "他们会在不同领域和训练阶段开展大规模同策略蒸馏", "selected_text": "他们会在不同领域和训练阶段开展大规模同策略蒸馏", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 635 ], "packet_id": 49, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1361.8712932692308, "end": 1365.731293269231, "text": "将所有专家教师模型的能力整合到同一个模型中", "raw_asr_text": "将所有专家教师模型的能力整合到同一个模型中", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 49, "words": [ { "word": "将", "start": 1361.8712932692308, "end": 1362.411293269231, "probability": 0.98779296875 }, { "word": "所有", "start": 1362.411293269231, "end": 1362.8312932692309, "probability": 0.99951171875 }, { "word": "专", "start": 1362.8312932692309, "end": 1363.0712932692309, "probability": 0.99951171875 }, { "word": "家", "start": 1363.0712932692309, "end": 1363.271293269231, "probability": 0.99951171875 }, { "word": "教", "start": 1363.271293269231, "end": 1363.471293269231, "probability": 0.998046875 }, { "word": "师", "start": 1363.471293269231, "end": 1363.6312932692308, "probability": 0.95068359375 }, { "word": "模", "start": 1363.6312932692308, "end": 1363.771293269231, "probability": 0.99951171875 }, { "word": "型", "start": 1363.771293269231, "end": 1363.931293269231, "probability": 0.99951171875 }, { "word": "的", "start": 1363.931293269231, "end": 1364.031293269231, "probability": 0.99951171875 }, { "word": "能", "start": 1364.031293269231, "end": 1364.171293269231, "probability": 0.99951171875 }, { "word": "力", "start": 1364.171293269231, "end": 1364.3712932692308, "probability": 0.99853515625 }, { "word": "整", "start": 1364.3712932692308, "end": 1364.5712932692309, "probability": 0.9990234375 }, { "word": "合", "start": 1364.5712932692309, "end": 1364.711293269231, "probability": 0.99951171875 }, { "word": "到", "start": 1364.711293269231, "end": 1364.8512932692308, "probability": 0.9990234375 }, { "word": "同", "start": 1364.8512932692308, "end": 1365.0512932692309, "probability": 0.99951171875 }, { "word": "一个", "start": 1365.0512932692309, "end": 1365.271293269231, "probability": 0.99853515625 }, { "word": "模", "start": 1365.271293269231, "end": 1365.411293269231, "probability": 0.99951171875 }, { "word": "型", "start": 1365.411293269231, "end": 1365.5712932692309, "probability": 1.0 }, { "word": "中", "start": 1365.5712932692309, "end": 1365.731293269231, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.0805326558775821, "compression_ratio": 1.2604790419161678, "no_speech_prob": 2.169375790117556e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "也就是说模型合并是用来持续汇总各次成功强化学习运行带来的提升等到流程最后他们会在不同领域和训练阶段开展大规模同策略蒸馏将所有专家教师模型的能力整合到同一个模型中另一方面推理力度也有一项升级它不再像V4那样只设置Think High和Think Max模式而是通过强化学习让推理力度成为一个从1到100的数值力度越低Token" } ], "id": 636, "review_annotation": "", "row_id": 636, "bounded_row_id": 636, "parent_row_ids": [ 636 ], "parent_row_id": 636, "packet_id": 49, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1351.011293269231, "packet_end": 1380.1992788461541, "raw_text": "将所有专家教师模型的能力整合到同一个模型中", "selected_text": "将所有专家教师模型的能力整合到同一个模型中", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 636 ], "packet_id": 49, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1365.731293269231, "end": 1368.391293269231, "text": "另一方面推理力度也有一项升级", "raw_asr_text": "另一方面推理力度也有一项升级", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 49, "words": [ { "word": "另", "start": 1365.731293269231, "end": 1366.0712932692309, "probability": 0.9951171875 }, { "word": "一", "start": 1366.0712932692309, "end": 1366.231293269231, "probability": 0.99755859375 }, { "word": "方面", "start": 1366.231293269231, "end": 1366.531293269231, "probability": 0.99951171875 }, { "word": "推", "start": 1366.531293269231, "end": 1366.8712932692308, "probability": 0.98876953125 }, { "word": "理", "start": 1366.8712932692308, "end": 1367.0712932692309, "probability": 1.0 }, { "word": "力", "start": 1367.0712932692309, "end": 1367.211293269231, "probability": 0.998046875 }, { "word": "度", "start": 1367.211293269231, "end": 1367.391293269231, "probability": 1.0 }, { "word": "也", "start": 1367.391293269231, "end": 1367.6112932692308, "probability": 0.99951171875 }, { "word": "有一", "start": 1367.6112932692308, "end": 1367.8312932692309, "probability": 0.99658203125 }, { "word": "项", "start": 1367.8312932692309, "end": 1367.971293269231, "probability": 0.998046875 }, { "word": "升", "start": 1367.971293269231, "end": 1368.151293269231, "probability": 0.9990234375 }, { "word": "级", "start": 1368.151293269231, "end": 1368.391293269231, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.0805326558775821, "compression_ratio": 1.2604790419161678, "no_speech_prob": 2.169375790117556e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "也就是说模型合并是用来持续汇总各次成功强化学习运行带来的提升等到流程最后他们会在不同领域和训练阶段开展大规模同策略蒸馏将所有专家教师模型的能力整合到同一个模型中另一方面推理力度也有一项升级它不再像V4那样只设置Think High和Think Max模式而是通过强化学习让推理力度成为一个从1到100的数值力度越低Token" } ], "id": 637, "review_annotation": "", "row_id": 637, "bounded_row_id": 637, "parent_row_ids": [ 637 ], "parent_row_id": 637, "packet_id": 49, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1351.011293269231, "packet_end": 1380.1992788461541, "raw_text": "另一方面推理力度也有一项升级", "selected_text": "另一方面推理力度也有一项升级", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 637 ], "packet_id": 49, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1368.391293269231, "end": 1373.171293269231, "text": "它不再像V4那样只设置Think High和Think Max模式", "raw_asr_text": "它不再像V4那样只设置Think High和Think Max模式", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 49, "words": [ { "word": "它", "start": 1368.391293269231, "end": 1368.8312932692309, "probability": 0.939453125 }, { "word": "不", "start": 1368.8312932692309, "end": 1369.1312932692308, "probability": 0.99951171875 }, { "word": "再", "start": 1369.1312932692308, "end": 1369.3512932692308, "probability": 0.9951171875 }, { "word": "像", "start": 1369.3512932692308, "end": 1369.5512932692309, "probability": 0.99951171875 }, { "word": "V", "start": 1369.5512932692309, "end": 1369.751293269231, "probability": 0.91357421875 }, { "word": "4", "start": 1369.751293269231, "end": 1369.951293269231, "probability": 0.96826171875 }, { "word": "那", "start": 1369.951293269231, "end": 1370.0912932692308, "probability": 0.99853515625 }, { "word": "样", "start": 1370.0912932692308, "end": 1370.271293269231, "probability": 0.99951171875 }, { "word": "只", "start": 1370.271293269231, "end": 1370.5512932692309, "probability": 0.9951171875 }, { "word": "设", "start": 1370.5512932692309, "end": 1370.8712932692308, "probability": 1.0 }, { "word": "置", "start": 1370.8712932692308, "end": 1371.0912932692308, "probability": 0.99462890625 }, { "word": "Th", "start": 1371.0912932692308, "end": 1371.391293269231, "probability": 0.88330078125 }, { "word": "ink", "start": 1371.391293269231, "end": 1371.531293269231, "probability": 0.9951171875 }, { "word": " High", "start": 1371.531293269231, "end": 1371.891293269231, "probability": 0.736328125 }, { "word": "和", "start": 1371.891293269231, "end": 1372.151293269231, "probability": 0.98779296875 }, { "word": "Th", "start": 1372.151293269231, "end": 1372.3112932692309, "probability": 0.99853515625 }, { "word": "ink", "start": 1372.3112932692309, "end": 1372.431293269231, "probability": 1.0 }, { "word": " Max", "start": 1372.431293269231, "end": 1372.731293269231, "probability": 0.8994140625 }, { "word": "模", "start": 1372.731293269231, "end": 1372.971293269231, "probability": 0.99658203125 }, { "word": "式", "start": 1372.971293269231, "end": 1373.171293269231, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.0805326558775821, "compression_ratio": 1.2604790419161678, "no_speech_prob": 2.169375790117556e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "也就是说模型合并是用来持续汇总各次成功强化学习运行带来的提升等到流程最后他们会在不同领域和训练阶段开展大规模同策略蒸馏将所有专家教师模型的能力整合到同一个模型中另一方面推理力度也有一项升级它不再像V4那样只设置Think High和Think Max模式而是通过强化学习让推理力度成为一个从1到100的数值力度越低Token" } ], "id": 638, "review_annotation": "", "row_id": 638, "bounded_row_id": 638, "parent_row_ids": [ 638 ], "parent_row_id": 638, "packet_id": 49, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1351.011293269231, "packet_end": 1380.1992788461541, "raw_text": "它不再像V4那样只设置Think High和Think Max模式", "selected_text": "它不再像V4那样只设置Think High和Think Max模式", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 638 ], "packet_id": 49, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1373.171293269231, "end": 1374.731293269231, "text": "而是通过强化学习", "raw_asr_text": "而是通过强化学习", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 49, "words": [ { "word": "而", "start": 1373.171293269231, "end": 1373.3712932692308, "probability": 0.9951171875 }, { "word": "是", "start": 1373.3712932692308, "end": 1373.491293269231, "probability": 1.0 }, { "word": "通", "start": 1373.491293269231, "end": 1373.6312932692308, "probability": 0.998046875 }, { "word": "过", "start": 1373.6312932692308, "end": 1373.8112932692309, "probability": 1.0 }, { "word": "强", "start": 1373.8112932692309, "end": 1374.1312932692308, "probability": 0.99951171875 }, { "word": "化", "start": 1374.1312932692308, "end": 1374.3112932692309, "probability": 1.0 }, { "word": "学", "start": 1374.3112932692309, "end": 1374.511293269231, "probability": 0.99951171875 }, { "word": "习", "start": 1374.511293269231, "end": 1374.731293269231, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.0805326558775821, "compression_ratio": 1.2604790419161678, "no_speech_prob": 2.169375790117556e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "也就是说模型合并是用来持续汇总各次成功强化学习运行带来的提升等到流程最后他们会在不同领域和训练阶段开展大规模同策略蒸馏将所有专家教师模型的能力整合到同一个模型中另一方面推理力度也有一项升级它不再像V4那样只设置Think High和Think Max模式而是通过强化学习让推理力度成为一个从1到100的数值力度越低Token" } ], "id": 639, "review_annotation": "", "row_id": 639, "bounded_row_id": 639, "parent_row_ids": [ 639 ], "parent_row_id": 639, "packet_id": 49, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1351.011293269231, "packet_end": 1380.1992788461541, "raw_text": "而是通过强化学习", "selected_text": "而是通过强化学习", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 639 ], "packet_id": 49, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1374.731293269231, "end": 1378.651293269231, "text": "让推理力度成为一个从1到100的数值", "raw_asr_text": "让推理力度成为一个从1到100的数值", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 49, "words": [ { "word": "让", "start": 1374.731293269231, "end": 1374.991293269231, "probability": 0.99072265625 }, { "word": "推", "start": 1374.991293269231, "end": 1375.211293269231, "probability": 0.99951171875 }, { "word": "理", "start": 1375.211293269231, "end": 1375.3712932692308, "probability": 1.0 }, { "word": "力", "start": 1375.3712932692308, "end": 1375.491293269231, "probability": 0.99853515625 }, { "word": "度", "start": 1375.491293269231, "end": 1375.671293269231, "probability": 0.99951171875 }, { "word": "成", "start": 1375.671293269231, "end": 1375.8512932692308, "probability": 0.9990234375 }, { "word": "为", "start": 1375.8512932692308, "end": 1376.031293269231, "probability": 0.99951171875 }, { "word": "一个", "start": 1376.031293269231, "end": 1376.211293269231, "probability": 0.994140625 }, { "word": "从", "start": 1376.211293269231, "end": 1376.531293269231, "probability": 0.9951171875 }, { "word": "1", "start": 1376.531293269231, "end": 1376.891293269231, "probability": 0.84033203125 }, { "word": "到", "start": 1376.891293269231, "end": 1377.0912932692308, "probability": 0.91552734375 }, { "word": "100", "start": 1377.0912932692308, "end": 1377.5512932692309, "probability": 0.99755859375 }, { "word": "的", "start": 1377.5512932692309, "end": 1377.971293269231, "probability": 0.99609375 }, { "word": "数", "start": 1377.971293269231, "end": 1378.3312932692309, "probability": 0.96435546875 }, { "word": "值", "start": 1378.3312932692309, "end": 1378.651293269231, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.0805326558775821, "compression_ratio": 1.2604790419161678, "no_speech_prob": 2.169375790117556e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "也就是说模型合并是用来持续汇总各次成功强化学习运行带来的提升等到流程最后他们会在不同领域和训练阶段开展大规模同策略蒸馏将所有专家教师模型的能力整合到同一个模型中另一方面推理力度也有一项升级它不再像V4那样只设置Think High和Think Max模式而是通过强化学习让推理力度成为一个从1到100的数值力度越低Token" } ], "id": 640, "review_annotation": "", "row_id": 640, "bounded_row_id": 640, "parent_row_ids": [ 640 ], "parent_row_id": 640, "packet_id": 49, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1351.011293269231, "packet_end": 1380.1992788461541, "raw_text": "让推理力度成为一个从1到100的数值", "selected_text": "让推理力度成为一个从1到100的数值", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 640 ], "packet_id": 49, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1378.651293269231, "end": 1379.3992788461542, "text": "力度越低", "raw_asr_text": "力度越低", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 49, "words": [ { "word": "力", "start": 1378.651293269231, "end": 1379.151293269231, "probability": 0.95751953125 }, { "word": "度", "start": 1379.151293269231, "end": 1379.3312932692309, "probability": 1.0 }, { "word": "越", "start": 1379.3312932692309, "end": 1379.511293269231, "probability": 0.9912109375 }, { "word": "低", "start": 1379.511293269231, "end": 1379.8112932692309, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.0805326558775821, "compression_ratio": 1.2604790419161678, "no_speech_prob": 2.169375790117556e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "也就是说模型合并是用来持续汇总各次成功强化学习运行带来的提升等到流程最后他们会在不同领域和训练阶段开展大规模同策略蒸馏将所有专家教师模型的能力整合到同一个模型中另一方面推理力度也有一项升级它不再像V4那样只设置Think High和Think Max模式而是通过强化学习让推理力度成为一个从1到100的数值力度越低Token" } ], "id": 641, "review_annotation": "", "row_id": 641, "bounded_row_id": 641, "parent_row_ids": [ 641 ], "parent_row_id": 641, "packet_id": 49, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1351.011293269231, "packet_end": 1380.1992788461541, "raw_text": "力度越低", "selected_text": "力度越低", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 641 ], "packet_id": 49, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1379.8192788461542, "end": 1380.9192788461542, "text": "Token惩罚越强", "raw_asr_text": "Token惩罚越强", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 50, "words": [ { "word": "T", "start": 1379.8192788461542, "end": 1380.0792788461542, "probability": 0.4072265625 }, { "word": "oken", "start": 1380.0792788461542, "end": 1380.219278846154, "probability": 0.99072265625 }, { "word": "惩", "start": 1380.219278846154, "end": 1380.3992788461542, "probability": 0.9775390625 }, { "word": "罚", "start": 1380.3992788461542, "end": 1380.5392788461543, "probability": 0.99755859375 }, { "word": "越", "start": 1380.5392788461543, "end": 1380.6592788461542, "probability": 0.98876953125 }, { "word": "强", "start": 1380.6592788461542, "end": 1380.9192788461542, "probability": 0.994140625 } ], "asr_metrics": { "avg_logprob": -0.10017968990184643, "compression_ratio": 1.2141057934508817, "no_speech_prob": 2.5706232281708097e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "力度越低Token惩罚越强力度越高就会逐步允许模型使用更多算力因此模型不是训练几个彼此孤立的推理模式而是学会一条平滑的成本质量曲线面向用户的Low High和Max 3档基本上就是同一尺度上力度值为50,75和100的设置好啦V4.1 Flash在技术报告中提出的大部分重要进展就介绍到这里这份51页的技术报告里还有许多技术细节但其中有些内容实在太深奥或者需要更多背景之外" } ], "id": 642, "review_annotation": "", "row_id": 642, "bounded_row_id": 642, "parent_row_ids": [ 642 ], "parent_row_id": 642, "packet_id": 50, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1378.5992788461542, "packet_end": 1407.787264423077, "raw_text": "Token惩罚越强", "selected_text": "Token惩罚越强", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 642 ], "packet_id": 50, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1380.9192788461542, "end": 1381.5992788461542, "text": "力度越高", "raw_asr_text": "力度越高", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 50, "words": [ { "word": "力", "start": 1380.9192788461542, "end": 1381.1192788461542, "probability": 0.98681640625 }, { "word": "度", "start": 1381.1192788461542, "end": 1381.2592788461543, "probability": 1.0 }, { "word": "越", "start": 1381.2592788461543, "end": 1381.3792788461542, "probability": 0.9990234375 }, { "word": "高", "start": 1381.3792788461542, "end": 1381.5992788461542, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.10017968990184643, "compression_ratio": 1.2141057934508817, "no_speech_prob": 2.5706232281708097e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "力度越低Token惩罚越强力度越高就会逐步允许模型使用更多算力因此模型不是训练几个彼此孤立的推理模式而是学会一条平滑的成本质量曲线面向用户的Low High和Max 3档基本上就是同一尺度上力度值为50,75和100的设置好啦V4.1 Flash在技术报告中提出的大部分重要进展就介绍到这里这份51页的技术报告里还有许多技术细节但其中有些内容实在太深奥或者需要更多背景之外" } ], "id": 643, "review_annotation": "", "row_id": 643, "bounded_row_id": 643, "parent_row_ids": [ 643 ], "parent_row_id": 643, "packet_id": 50, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1378.5992788461542, "packet_end": 1407.787264423077, "raw_text": "力度越高", "selected_text": "力度越高", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 643 ], "packet_id": 50, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1381.5992788461542, "end": 1384.1592788461542, "text": "就会逐步允许模型使用更多算力", "raw_asr_text": "就会逐步允许模型使用更多算力", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 50, "words": [ { "word": "就", "start": 1381.5992788461542, "end": 1381.7792788461543, "probability": 0.98486328125 }, { "word": "会", "start": 1381.7792788461543, "end": 1381.9192788461542, "probability": 0.9990234375 }, { "word": "逐", "start": 1381.9192788461542, "end": 1382.1592788461542, "probability": 0.9990234375 }, { "word": "步", "start": 1382.1592788461542, "end": 1382.3192788461542, "probability": 0.99853515625 }, { "word": "允", "start": 1382.3192788461542, "end": 1382.5592788461543, "probability": 0.9990234375 }, { "word": "许", "start": 1382.5592788461543, "end": 1382.719278846154, "probability": 0.9990234375 }, { "word": "模", "start": 1382.719278846154, "end": 1382.9192788461542, "probability": 0.99560546875 }, { "word": "型", "start": 1382.9192788461542, "end": 1383.0992788461542, "probability": 0.99853515625 }, { "word": "使", "start": 1383.0992788461542, "end": 1383.2792788461543, "probability": 0.9990234375 }, { "word": "用", "start": 1383.2792788461543, "end": 1383.4192788461542, "probability": 1.0 }, { "word": "更", "start": 1383.4192788461542, "end": 1383.5992788461542, "probability": 0.99853515625 }, { "word": "多", "start": 1383.5992788461542, "end": 1383.7792788461543, "probability": 1.0 }, { "word": "算", "start": 1383.7792788461543, "end": 1383.9592788461541, "probability": 0.9970703125 }, { "word": "力", "start": 1383.9592788461541, "end": 1384.1592788461542, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.10017968990184643, "compression_ratio": 1.2141057934508817, "no_speech_prob": 2.5706232281708097e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "力度越低Token惩罚越强力度越高就会逐步允许模型使用更多算力因此模型不是训练几个彼此孤立的推理模式而是学会一条平滑的成本质量曲线面向用户的Low High和Max 3档基本上就是同一尺度上力度值为50,75和100的设置好啦V4.1 Flash在技术报告中提出的大部分重要进展就介绍到这里这份51页的技术报告里还有许多技术细节但其中有些内容实在太深奥或者需要更多背景之外" } ], "id": 644, "review_annotation": "", "row_id": 644, "bounded_row_id": 644, "parent_row_ids": [ 644 ], "parent_row_id": 644, "packet_id": 50, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1378.5992788461542, "packet_end": 1407.787264423077, "raw_text": "就会逐步允许模型使用更多算力", "selected_text": "就会逐步允许模型使用更多算力", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 644 ], "packet_id": 50, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1384.1592788461542, "end": 1384.8392788461542, "text": "因此", "raw_asr_text": "因此", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 50, "words": [ { "word": "因", "start": 1384.1592788461542, "end": 1384.6192788461542, "probability": 0.99755859375 }, { "word": "此", "start": 1384.6192788461542, "end": 1384.8392788461542, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10017968990184643, "compression_ratio": 1.2141057934508817, "no_speech_prob": 2.5706232281708097e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "力度越低Token惩罚越强力度越高就会逐步允许模型使用更多算力因此模型不是训练几个彼此孤立的推理模式而是学会一条平滑的成本质量曲线面向用户的Low High和Max 3档基本上就是同一尺度上力度值为50,75和100的设置好啦V4.1 Flash在技术报告中提出的大部分重要进展就介绍到这里这份51页的技术报告里还有许多技术细节但其中有些内容实在太深奥或者需要更多背景之外" } ], "id": 645, "review_annotation": "", "row_id": 645, "bounded_row_id": 645, "parent_row_ids": [ 645 ], "parent_row_id": 645, "packet_id": 50, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1378.5992788461542, "packet_end": 1407.787264423077, "raw_text": "因此", "selected_text": "因此", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 645 ], "packet_id": 50, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1384.8392788461542, "end": 1387.2992788461543, "text": "模型不是训练几个彼此孤立的推理模式", "raw_asr_text": "模型不是训练几个彼此孤立的推理模式", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 50, "words": [ { "word": "模", "start": 1384.8392788461542, "end": 1385.1592788461542, "probability": 0.919921875 }, { "word": "型", "start": 1385.1592788461542, "end": 1385.3192788461542, "probability": 1.0 }, { "word": "不是", "start": 1385.3192788461542, "end": 1385.5592788461543, "probability": 0.99951171875 }, { "word": "训", "start": 1385.5592788461543, "end": 1385.719278846154, "probability": 0.9990234375 }, { "word": "练", "start": 1385.719278846154, "end": 1385.8392788461542, "probability": 0.99951171875 }, { "word": "几", "start": 1385.8392788461542, "end": 1385.9992788461543, "probability": 0.998046875 }, { "word": "个", "start": 1385.9992788461543, "end": 1386.0992788461542, "probability": 1.0 }, { "word": "彼", "start": 1386.0992788461542, "end": 1386.2592788461543, "probability": 0.9990234375 }, { "word": "此", "start": 1386.2592788461543, "end": 1386.3792788461542, "probability": 1.0 }, { "word": "孤", "start": 1386.3792788461542, "end": 1386.5592788461543, "probability": 0.9990234375 }, { "word": "立", "start": 1386.5592788461543, "end": 1386.6392788461542, "probability": 0.9873046875 }, { "word": "的", "start": 1386.6392788461542, "end": 1386.7392788461543, "probability": 0.9990234375 }, { "word": "推", "start": 1386.7392788461543, "end": 1386.8592788461542, "probability": 1.0 }, { "word": "理", "start": 1386.8592788461542, "end": 1386.9592788461541, "probability": 0.99951171875 }, { "word": "模", "start": 1386.9592788461541, "end": 1387.0792788461542, "probability": 0.99951171875 }, { "word": "式", "start": 1387.0792788461542, "end": 1387.2992788461543, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.10017968990184643, "compression_ratio": 1.2141057934508817, "no_speech_prob": 2.5706232281708097e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "力度越低Token惩罚越强力度越高就会逐步允许模型使用更多算力因此模型不是训练几个彼此孤立的推理模式而是学会一条平滑的成本质量曲线面向用户的Low High和Max 3档基本上就是同一尺度上力度值为50,75和100的设置好啦V4.1 Flash在技术报告中提出的大部分重要进展就介绍到这里这份51页的技术报告里还有许多技术细节但其中有些内容实在太深奥或者需要更多背景之外" } ], "id": 646, "review_annotation": "", "row_id": 646, "bounded_row_id": 646, "parent_row_ids": [ 646 ], "parent_row_id": 646, "packet_id": 50, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1378.5992788461542, "packet_end": 1407.787264423077, "raw_text": "模型不是训练几个彼此孤立的推理模式", "selected_text": "模型不是训练几个彼此孤立的推理模式", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 646 ], "packet_id": 50, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1387.2992788461543, "end": 1389.6192788461542, "text": "而是学会一条平滑的成本质量曲线", "raw_asr_text": "而是学会一条平滑的成本质量曲线", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 50, "words": [ { "word": "而", "start": 1387.2992788461543, "end": 1387.5392788461543, "probability": 0.99853515625 }, { "word": "是", "start": 1387.5392788461543, "end": 1387.6592788461542, "probability": 0.99951171875 }, { "word": "学", "start": 1387.6592788461542, "end": 1387.7592788461543, "probability": 0.99853515625 }, { "word": "会", "start": 1387.7592788461543, "end": 1387.8792788461542, "probability": 0.98779296875 }, { "word": "一", "start": 1387.8792788461542, "end": 1387.9592788461541, "probability": 0.99609375 }, { "word": "条", "start": 1387.9592788461541, "end": 1388.0792788461542, "probability": 0.99853515625 }, { "word": "平", "start": 1388.0792788461542, "end": 1388.2992788461543, "probability": 0.99951171875 }, { "word": "滑", "start": 1388.2992788461543, "end": 1388.5392788461543, "probability": 0.99853515625 }, { "word": "的", "start": 1388.5392788461543, "end": 1388.6592788461542, "probability": 0.99755859375 }, { "word": "成", "start": 1388.6592788461542, "end": 1388.8392788461542, "probability": 0.99951171875 }, { "word": "本", "start": 1388.8392788461542, "end": 1388.9592788461541, "probability": 0.9990234375 }, { "word": "质", "start": 1388.9592788461541, "end": 1389.0992788461542, "probability": 0.9990234375 }, { "word": "量", "start": 1389.0992788461542, "end": 1389.1992788461541, "probability": 1.0 }, { "word": "曲", "start": 1389.1992788461541, "end": 1389.3792788461542, "probability": 0.99169921875 }, { "word": "线", "start": 1389.3792788461542, "end": 1389.6192788461542, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.10017968990184643, "compression_ratio": 1.2141057934508817, "no_speech_prob": 2.5706232281708097e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "力度越低Token惩罚越强力度越高就会逐步允许模型使用更多算力因此模型不是训练几个彼此孤立的推理模式而是学会一条平滑的成本质量曲线面向用户的Low High和Max 3档基本上就是同一尺度上力度值为50,75和100的设置好啦V4.1 Flash在技术报告中提出的大部分重要进展就介绍到这里这份51页的技术报告里还有许多技术细节但其中有些内容实在太深奥或者需要更多背景之外" } ], "id": 647, "review_annotation": "", "row_id": 647, "bounded_row_id": 647, "parent_row_ids": [ 647 ], "parent_row_id": 647, "packet_id": 50, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1378.5992788461542, "packet_end": 1407.787264423077, "raw_text": "而是学会一条平滑的成本质量曲线", "selected_text": "而是学会一条平滑的成本质量曲线", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 647 ], "packet_id": 50, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1389.6192788461542, "end": 1392.0392788461543, "text": "面向用户的Low High和Max 3档", "raw_asr_text": "面向用户的Low High和Max 3档", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 50, "words": [ { "word": "面", "start": 1389.6192788461542, "end": 1389.8192788461542, "probability": 0.98828125 }, { "word": "向", "start": 1389.8192788461542, "end": 1389.9592788461541, "probability": 0.978515625 }, { "word": "用", "start": 1389.9592788461541, "end": 1390.1992788461541, "probability": 0.998046875 }, { "word": "户", "start": 1390.1992788461541, "end": 1390.3392788461542, "probability": 0.99951171875 }, { "word": "的", "start": 1390.3392788461542, "end": 1390.4592788461541, "probability": 0.9970703125 }, { "word": "L", "start": 1390.4592788461541, "end": 1390.6792788461541, "probability": 0.65771484375 }, { "word": "ow", "start": 1390.6792788461541, "end": 1390.7592788461543, "probability": 0.96337890625 }, { "word": " High", "start": 1390.7592788461543, "end": 1390.9592788461541, "probability": 0.6787109375 }, { "word": "和", "start": 1390.9592788461541, "end": 1391.2592788461543, "probability": 0.9921875 }, { "word": "Max", "start": 1391.2592788461543, "end": 1391.5392788461543, "probability": 0.9326171875 }, { "word": " 3", "start": 1391.5392788461543, "end": 1391.8392788461542, "probability": 0.380859375 }, { "word": "档", "start": 1391.8392788461542, "end": 1392.0392788461543, "probability": 0.87109375 } ], "asr_metrics": { "avg_logprob": -0.10017968990184643, "compression_ratio": 1.2141057934508817, "no_speech_prob": 2.5706232281708097e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "力度越低Token惩罚越强力度越高就会逐步允许模型使用更多算力因此模型不是训练几个彼此孤立的推理模式而是学会一条平滑的成本质量曲线面向用户的Low High和Max 3档基本上就是同一尺度上力度值为50,75和100的设置好啦V4.1 Flash在技术报告中提出的大部分重要进展就介绍到这里这份51页的技术报告里还有许多技术细节但其中有些内容实在太深奥或者需要更多背景之外" } ], "id": 648, "review_annotation": "", "row_id": 648, "bounded_row_id": 648, "parent_row_ids": [ 648 ], "parent_row_id": 648, "packet_id": 50, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1378.5992788461542, "packet_end": 1407.787264423077, "raw_text": "面向用户的Low High和Max 3档", "selected_text": "面向用户的Low High和Max 3档", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 648 ], "packet_id": 50, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1392.0392788461543, "end": 1393.8792788461542, "text": "基本上就是同一尺度上", "raw_asr_text": "基本上就是同一尺度上", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 50, "words": [ { "word": "基本", "start": 1392.0392788461543, "end": 1392.4592788461541, "probability": 0.99755859375 }, { "word": "上", "start": 1392.4592788461541, "end": 1392.6792788461541, "probability": 0.99365234375 }, { "word": "就是", "start": 1392.6792788461541, "end": 1392.9392788461541, "probability": 0.990234375 }, { "word": "同", "start": 1392.9392788461541, "end": 1393.2392788461543, "probability": 0.98828125 }, { "word": "一", "start": 1393.2392788461543, "end": 1393.4192788461542, "probability": 0.93212890625 }, { "word": "尺", "start": 1393.4192788461542, "end": 1393.5592788461543, "probability": 0.99951171875 }, { "word": "度", "start": 1393.5592788461543, "end": 1393.6992788461541, "probability": 0.99951171875 }, { "word": "上", "start": 1393.6992788461541, "end": 1393.8792788461542, "probability": 0.99462890625 } ], "asr_metrics": { "avg_logprob": -0.10017968990184643, "compression_ratio": 1.2141057934508817, "no_speech_prob": 2.5706232281708097e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "力度越低Token惩罚越强力度越高就会逐步允许模型使用更多算力因此模型不是训练几个彼此孤立的推理模式而是学会一条平滑的成本质量曲线面向用户的Low High和Max 3档基本上就是同一尺度上力度值为50,75和100的设置好啦V4.1 Flash在技术报告中提出的大部分重要进展就介绍到这里这份51页的技术报告里还有许多技术细节但其中有些内容实在太深奥或者需要更多背景之外" } ], "id": 649, "review_annotation": "", "row_id": 649, "bounded_row_id": 649, "parent_row_ids": [ 649 ], "parent_row_id": 649, "packet_id": 50, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1378.5992788461542, "packet_end": 1407.787264423077, "raw_text": "基本上就是同一尺度上", "selected_text": "基本上就是同一尺度上", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 649 ], "packet_id": 50, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1393.8792788461542, "end": 1396.4792788461543, "text": "力度值为50,75和100的设置", "raw_asr_text": "力度值为50,75和100的设置", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 50, "words": [ { "word": "力", "start": 1393.8792788461542, "end": 1394.0792788461542, "probability": 0.99462890625 }, { "word": "度", "start": 1394.0792788461542, "end": 1394.219278846154, "probability": 0.99853515625 }, { "word": "值", "start": 1394.219278846154, "end": 1394.3592788461542, "probability": 0.96630859375 }, { "word": "为", "start": 1394.3592788461542, "end": 1394.5792788461542, "probability": 0.9921875 }, { "word": "50,", "start": 1394.5792788461542, "end": 1394.9192788461542, "probability": 0.9033203125 }, { "word": "75", "start": 1395.0592788461543, "end": 1395.3392788461542, "probability": 0.974609375 }, { "word": "和", "start": 1395.3392788461542, "end": 1395.6992788461541, "probability": 0.98583984375 }, { "word": "100", "start": 1395.6992788461541, "end": 1395.9992788461543, "probability": 0.99853515625 }, { "word": "的", "start": 1395.9992788461543, "end": 1396.1592788461542, "probability": 0.9970703125 }, { "word": "设", "start": 1396.1592788461542, "end": 1396.3192788461542, "probability": 0.9990234375 }, { "word": "置", "start": 1396.3192788461542, "end": 1396.4792788461543, "probability": 0.99853515625 } ], "asr_metrics": { "avg_logprob": -0.10017968990184643, "compression_ratio": 1.2141057934508817, "no_speech_prob": 2.5706232281708097e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "力度越低Token惩罚越强力度越高就会逐步允许模型使用更多算力因此模型不是训练几个彼此孤立的推理模式而是学会一条平滑的成本质量曲线面向用户的Low High和Max 3档基本上就是同一尺度上力度值为50,75和100的设置好啦V4.1 Flash在技术报告中提出的大部分重要进展就介绍到这里这份51页的技术报告里还有许多技术细节但其中有些内容实在太深奥或者需要更多背景之外" } ], "id": 650, "review_annotation": "", "row_id": 650, "bounded_row_id": 650, "parent_row_ids": [ 650 ], "parent_row_id": 650, "packet_id": 50, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1378.5992788461542, "packet_end": 1407.787264423077, "raw_text": "力度值为50,75和100的设置", "selected_text": "力度值为50,75和100的设置", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 650 ], "packet_id": 50, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1396.4792788461543, "end": 1397.0392788461543, "text": "好啦", "raw_asr_text": "好啦", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 50, "words": [ { "word": "好啦", "start": 1396.4792788461543, "end": 1397.0392788461543, "probability": 0.37109375 } ], "asr_metrics": { "avg_logprob": -0.10017968990184643, "compression_ratio": 1.2141057934508817, "no_speech_prob": 2.5706232281708097e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "力度越低Token惩罚越强力度越高就会逐步允许模型使用更多算力因此模型不是训练几个彼此孤立的推理模式而是学会一条平滑的成本质量曲线面向用户的Low High和Max 3档基本上就是同一尺度上力度值为50,75和100的设置好啦V4.1 Flash在技术报告中提出的大部分重要进展就介绍到这里这份51页的技术报告里还有许多技术细节但其中有些内容实在太深奥或者需要更多背景之外" } ], "id": 651, "review_annotation": "", "row_id": 651, "bounded_row_id": 651, "parent_row_ids": [ 651 ], "parent_row_id": 651, "packet_id": 50, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1378.5992788461542, "packet_end": 1407.787264423077, "raw_text": "好啦", "selected_text": "好啦", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 651 ], "packet_id": 50, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1397.0392788461543, "end": 1399.3392788461542, "text": "V4.1 Flash在技术报告中提出的", "raw_asr_text": "V4.1 Flash在技术报告中提出的", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 50, "words": [ { "word": "V", "start": 1397.0392788461543, "end": 1397.3192788461542, "probability": 0.495849609375 }, { "word": "4.", "start": 1397.3192788461542, "end": 1397.4792788461543, "probability": 0.98681640625 }, { "word": "1", "start": 1397.5792788461542, "end": 1397.6992788461541, "probability": 0.98876953125 }, { "word": " Flash", "start": 1397.6992788461541, "end": 1397.9392788461541, "probability": 0.39404296875 }, { "word": "在", "start": 1397.9392788461541, "end": 1398.1792788461541, "probability": 0.9853515625 }, { "word": "技", "start": 1398.1792788461541, "end": 1398.3392788461542, "probability": 0.99169921875 }, { "word": "术", "start": 1398.3392788461542, "end": 1398.4592788461541, "probability": 0.99951171875 }, { "word": "报", "start": 1398.4592788461541, "end": 1398.5992788461542, "probability": 0.99951171875 }, { "word": "告", "start": 1398.5992788461542, "end": 1398.719278846154, "probability": 0.99755859375 }, { "word": "中", "start": 1398.719278846154, "end": 1398.8592788461542, "probability": 0.99658203125 }, { "word": "提", "start": 1398.8592788461542, "end": 1399.0392788461543, "probability": 0.99853515625 }, { "word": "出", "start": 1399.0392788461543, "end": 1399.219278846154, "probability": 0.9990234375 }, { "word": "的", "start": 1399.219278846154, "end": 1399.3392788461542, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.10017968990184643, "compression_ratio": 1.2141057934508817, "no_speech_prob": 2.5706232281708097e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "力度越低Token惩罚越强力度越高就会逐步允许模型使用更多算力因此模型不是训练几个彼此孤立的推理模式而是学会一条平滑的成本质量曲线面向用户的Low High和Max 3档基本上就是同一尺度上力度值为50,75和100的设置好啦V4.1 Flash在技术报告中提出的大部分重要进展就介绍到这里这份51页的技术报告里还有许多技术细节但其中有些内容实在太深奥或者需要更多背景之外" } ], "id": 652, "review_annotation": "", "row_id": 652, "bounded_row_id": 652, "parent_row_ids": [ 652 ], "parent_row_id": 652, "packet_id": 50, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1378.5992788461542, "packet_end": 1407.787264423077, "raw_text": "V4.1 Flash在技术报告中提出的", "selected_text": "V4.1 Flash在技术报告中提出的", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 652 ], "packet_id": 50, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1399.3392788461542, "end": 1401.0992788461542, "text": "大部分重要进展就介绍到这里", "raw_asr_text": "大部分重要进展就介绍到这里", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 50, "words": [ { "word": "大", "start": 1399.3392788461542, "end": 1399.4792788461543, "probability": 0.998046875 }, { "word": "部分", "start": 1399.4792788461543, "end": 1399.6792788461541, "probability": 0.99755859375 }, { "word": "重要", "start": 1399.6792788461541, "end": 1399.8992788461542, "probability": 0.98974609375 }, { "word": "进", "start": 1399.8992788461542, "end": 1400.0592788461543, "probability": 0.9951171875 }, { "word": "展", "start": 1400.0592788461543, "end": 1400.2392788461543, "probability": 0.99951171875 }, { "word": "就", "start": 1400.2392788461543, "end": 1400.4592788461541, "probability": 0.99609375 }, { "word": "介", "start": 1400.4592788461541, "end": 1400.5992788461542, "probability": 0.9990234375 }, { "word": "绍", "start": 1400.5992788461542, "end": 1400.7392788461543, "probability": 1.0 }, { "word": "到", "start": 1400.7392788461543, "end": 1400.8392788461542, "probability": 0.99853515625 }, { "word": "这里", "start": 1400.8392788461542, "end": 1401.0992788461542, "probability": 0.99755859375 } ], "asr_metrics": { "avg_logprob": -0.10017968990184643, "compression_ratio": 1.2141057934508817, "no_speech_prob": 2.5706232281708097e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "力度越低Token惩罚越强力度越高就会逐步允许模型使用更多算力因此模型不是训练几个彼此孤立的推理模式而是学会一条平滑的成本质量曲线面向用户的Low High和Max 3档基本上就是同一尺度上力度值为50,75和100的设置好啦V4.1 Flash在技术报告中提出的大部分重要进展就介绍到这里这份51页的技术报告里还有许多技术细节但其中有些内容实在太深奥或者需要更多背景之外" } ], "id": 653, "review_annotation": "", "row_id": 653, "bounded_row_id": 653, "parent_row_ids": [ 653 ], "parent_row_id": 653, "packet_id": 50, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1378.5992788461542, "packet_end": 1407.787264423077, "raw_text": "大部分重要进展就介绍到这里", "selected_text": "大部分重要进展就介绍到这里", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 653 ], "packet_id": 50, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1401.0992788461542, "end": 1402.9992788461543, "text": "这份51页的技术报告里", "raw_asr_text": "这份51页的技术报告里", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 50, "words": [ { "word": "这", "start": 1401.0992788461542, "end": 1401.4592788461541, "probability": 0.99853515625 }, { "word": "份", "start": 1401.4592788461541, "end": 1401.6192788461542, "probability": 0.99853515625 }, { "word": "51", "start": 1401.6192788461542, "end": 1402.0192788461543, "probability": 0.98583984375 }, { "word": "页", "start": 1402.0192788461543, "end": 1402.1992788461541, "probability": 0.998046875 }, { "word": "的", "start": 1402.1992788461541, "end": 1402.2992788461543, "probability": 0.9970703125 }, { "word": "技", "start": 1402.2992788461543, "end": 1402.4392788461541, "probability": 0.9931640625 }, { "word": "术", "start": 1402.4392788461541, "end": 1402.5592788461543, "probability": 1.0 }, { "word": "报", "start": 1402.5592788461543, "end": 1402.719278846154, "probability": 0.9990234375 }, { "word": "告", "start": 1402.719278846154, "end": 1402.9192788461542, "probability": 0.9970703125 }, { "word": "里", "start": 1402.9192788461542, "end": 1402.9992788461543, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.10017968990184643, "compression_ratio": 1.2141057934508817, "no_speech_prob": 2.5706232281708097e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "力度越低Token惩罚越强力度越高就会逐步允许模型使用更多算力因此模型不是训练几个彼此孤立的推理模式而是学会一条平滑的成本质量曲线面向用户的Low High和Max 3档基本上就是同一尺度上力度值为50,75和100的设置好啦V4.1 Flash在技术报告中提出的大部分重要进展就介绍到这里这份51页的技术报告里还有许多技术细节但其中有些内容实在太深奥或者需要更多背景之外" } ], "id": 654, "review_annotation": "", "row_id": 654, "bounded_row_id": 654, "parent_row_ids": [ 654 ], "parent_row_id": 654, "packet_id": 50, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1378.5992788461542, "packet_end": 1407.787264423077, "raw_text": "这份51页的技术报告里", "selected_text": "这份51页的技术报告里", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 654 ], "packet_id": 50, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1402.9992788461543, "end": 1404.2792788461543, "text": "还有许多技术细节", "raw_asr_text": "还有许多技术细节", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 50, "words": [ { "word": "还有", "start": 1402.9992788461543, "end": 1403.2792788461543, "probability": 0.99755859375 }, { "word": "许", "start": 1403.2792788461543, "end": 1403.5392788461543, "probability": 1.0 }, { "word": "多", "start": 1403.5392788461543, "end": 1403.6392788461542, "probability": 1.0 }, { "word": "技", "start": 1403.6392788461542, "end": 1403.8192788461542, "probability": 0.998046875 }, { "word": "术", "start": 1403.8192788461542, "end": 1403.9392788461541, "probability": 1.0 }, { "word": "细", "start": 1403.9392788461541, "end": 1404.1192788461542, "probability": 0.998046875 }, { "word": "节", "start": 1404.1192788461542, "end": 1404.2792788461543, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.10017968990184643, "compression_ratio": 1.2141057934508817, "no_speech_prob": 2.5706232281708097e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "力度越低Token惩罚越强力度越高就会逐步允许模型使用更多算力因此模型不是训练几个彼此孤立的推理模式而是学会一条平滑的成本质量曲线面向用户的Low High和Max 3档基本上就是同一尺度上力度值为50,75和100的设置好啦V4.1 Flash在技术报告中提出的大部分重要进展就介绍到这里这份51页的技术报告里还有许多技术细节但其中有些内容实在太深奥或者需要更多背景之外" } ], "id": 655, "review_annotation": "", "row_id": 655, "bounded_row_id": 655, "parent_row_ids": [ 655 ], "parent_row_id": 655, "packet_id": 50, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1378.5992788461542, "packet_end": 1407.787264423077, "raw_text": "还有许多技术细节", "selected_text": "还有许多技术细节", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 655 ], "packet_id": 50, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1404.2792788461543, "end": 1406.4192788461542, "text": "但其中有些内容实在太深奥", "raw_asr_text": "但其中有些内容实在太深奥", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 50, "words": [ { "word": "但", "start": 1404.2792788461543, "end": 1404.6992788461541, "probability": 0.99609375 }, { "word": "其", "start": 1404.6992788461541, "end": 1404.8792788461542, "probability": 0.9990234375 }, { "word": "中", "start": 1404.8792788461542, "end": 1405.0192788461543, "probability": 1.0 }, { "word": "有", "start": 1405.0192788461543, "end": 1405.1992788461541, "probability": 0.998046875 }, { "word": "些", "start": 1405.1992788461541, "end": 1405.3192788461542, "probability": 0.99951171875 }, { "word": "内", "start": 1405.3192788461542, "end": 1405.4592788461541, "probability": 0.998046875 }, { "word": "容", "start": 1405.4592788461541, "end": 1405.5992788461542, "probability": 1.0 }, { "word": "实", "start": 1405.5992788461542, "end": 1405.7992788461543, "probability": 0.9990234375 }, { "word": "在", "start": 1405.7992788461543, "end": 1405.9392788461541, "probability": 1.0 }, { "word": "太", "start": 1405.9392788461541, "end": 1406.0792788461542, "probability": 0.9970703125 }, { "word": "深", "start": 1406.0792788461542, "end": 1406.2392788461543, "probability": 1.0 }, { "word": "奥", "start": 1406.2392788461543, "end": 1406.4192788461542, "probability": 0.97607421875 } ], "asr_metrics": { "avg_logprob": -0.10017968990184643, "compression_ratio": 1.2141057934508817, "no_speech_prob": 2.5706232281708097e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "力度越低Token惩罚越强力度越高就会逐步允许模型使用更多算力因此模型不是训练几个彼此孤立的推理模式而是学会一条平滑的成本质量曲线面向用户的Low High和Max 3档基本上就是同一尺度上力度值为50,75和100的设置好啦V4.1 Flash在技术报告中提出的大部分重要进展就介绍到这里这份51页的技术报告里还有许多技术细节但其中有些内容实在太深奥或者需要更多背景之外" } ], "id": 656, "review_annotation": "", "row_id": 656, "bounded_row_id": 656, "parent_row_ids": [ 656 ], "parent_row_id": 656, "packet_id": 50, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1378.5992788461542, "packet_end": 1407.787264423077, "raw_text": "但其中有些内容实在太深奥", "selected_text": "但其中有些内容实在太深奥", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 656 ], "packet_id": 50, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1406.987264423077, "end": 1408.8272644230772, "text": "或者需要更多背景知识才能解释清楚", "raw_asr_text": "或者需要更多背景知识才能解释清楚", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 51, "words": [ { "word": "或者", "start": 1406.187264423077, "end": 1406.667264423077, "probability": 0.8642578125 }, { "word": "需要", "start": 1406.667264423077, "end": 1406.947264423077, "probability": 0.98388671875 }, { "word": "更", "start": 1406.947264423077, "end": 1407.167264423077, "probability": 0.99609375 }, { "word": "多", "start": 1407.167264423077, "end": 1407.3472644230771, "probability": 0.99853515625 }, { "word": "背", "start": 1407.3472644230771, "end": 1407.527264423077, "probability": 0.9326171875 }, { "word": "景", "start": 1407.527264423077, "end": 1407.687264423077, "probability": 1.0 }, { "word": "知", "start": 1407.687264423077, "end": 1407.8672644230771, "probability": 0.98095703125 }, { "word": "识", "start": 1407.8672644230771, "end": 1408.027264423077, "probability": 0.9296875 }, { "word": "才", "start": 1408.027264423077, "end": 1408.187264423077, "probability": 0.94970703125 }, { "word": "能", "start": 1408.187264423077, "end": 1408.307264423077, "probability": 0.9951171875 }, { "word": "解", "start": 1408.307264423077, "end": 1408.427264423077, "probability": 0.99853515625 }, { "word": "释", "start": 1408.427264423077, "end": 1408.5872644230772, "probability": 0.99658203125 }, { "word": "清楚", "start": 1408.5872644230772, "end": 1408.8272644230772, "probability": 0.99609375 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 657, "review_annotation": "", "row_id": 657, "bounded_row_id": 657, "parent_row_ids": [ 657 ], "parent_row_id": 657, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "或者需要更多背景知识才能解释清楚", "selected_text": "或者需要更多背景知识才能解释清楚", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 657 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1408.8272644230772, "end": 1412.0872644230772, "text": "比如DSEC缩放多莫态混合专家均衡", "raw_asr_text": "比如DSEC缩放多莫态混合专家均衡", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 51, "words": [ { "word": "比如", "start": 1408.8272644230772, "end": 1409.427264423077, "probability": 0.93994140625 }, { "word": "D", "start": 1409.427264423077, "end": 1409.707264423077, "probability": 0.39501953125 }, { "word": "SE", "start": 1409.707264423077, "end": 1409.907264423077, "probability": 0.469970703125 }, { "word": "C", "start": 1409.907264423077, "end": 1410.027264423077, "probability": 0.9052734375 }, { "word": "缩", "start": 1410.027264423077, "end": 1410.267264423077, "probability": 0.92578125 }, { "word": "放", "start": 1410.267264423077, "end": 1410.527264423077, "probability": 0.9912109375 }, { "word": "多", "start": 1410.527264423077, "end": 1410.747264423077, "probability": 0.72705078125 }, { "word": "莫", "start": 1410.747264423077, "end": 1410.907264423077, "probability": 0.72412109375 }, { "word": "态", "start": 1410.907264423077, "end": 1411.027264423077, "probability": 0.68408203125 }, { "word": "混", "start": 1411.027264423077, "end": 1411.167264423077, "probability": 0.99609375 }, { "word": "合", "start": 1411.167264423077, "end": 1411.307264423077, "probability": 0.9775390625 }, { "word": "专", "start": 1411.307264423077, "end": 1411.447264423077, "probability": 0.990234375 }, { "word": "家", "start": 1411.447264423077, "end": 1411.6272644230771, "probability": 0.99755859375 }, { "word": "均", "start": 1411.6272644230771, "end": 1411.887264423077, "probability": 0.96484375 }, { "word": "衡", "start": 1411.887264423077, "end": 1412.0872644230772, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 658, "review_annotation": "", "row_id": 658, "bounded_row_id": 658, "parent_row_ids": [ 658 ], "parent_row_id": 658, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "比如DSEC缩放多莫态混合专家均衡", "selected_text": "比如DSEC缩放多莫态混合专家均衡", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 658 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1412.0872644230772, "end": 1413.507264423077, "text": "以及其他基础设施改进", "raw_asr_text": "以及其他基础设施改进", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 51, "words": [ { "word": "以及", "start": 1412.0872644230772, "end": 1412.3272644230772, "probability": 0.9736328125 }, { "word": "其", "start": 1412.3272644230772, "end": 1412.5872644230772, "probability": 0.9931640625 }, { "word": "他", "start": 1412.5872644230772, "end": 1412.707264423077, "probability": 0.89990234375 }, { "word": "基", "start": 1412.707264423077, "end": 1412.8272644230772, "probability": 0.8359375 }, { "word": "础", "start": 1412.8272644230772, "end": 1412.947264423077, "probability": 0.99658203125 }, { "word": "设", "start": 1412.947264423077, "end": 1413.0672644230772, "probability": 0.990234375 }, { "word": "施", "start": 1413.0672644230772, "end": 1413.167264423077, "probability": 0.99609375 }, { "word": "改", "start": 1413.167264423077, "end": 1413.307264423077, "probability": 0.99853515625 }, { "word": "进", "start": 1413.307264423077, "end": 1413.507264423077, "probability": 0.99560546875 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 659, "review_annotation": "", "row_id": 659, "bounded_row_id": 659, "parent_row_ids": [ 659 ], "parent_row_id": 659, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "以及其他基础设施改进", "selected_text": "以及其他基础设施改进", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 659 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1413.507264423077, "end": 1415.007264423077, "text": "所以如果你们想了解更多", "raw_asr_text": "所以如果你们想了解更多", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 51, "words": [ { "word": "所以", "start": 1413.507264423077, "end": 1413.8472644230771, "probability": 0.9541015625 }, { "word": "如果", "start": 1413.8472644230771, "end": 1414.0672644230772, "probability": 0.9541015625 }, { "word": "你们", "start": 1414.0672644230772, "end": 1414.267264423077, "probability": 0.97412109375 }, { "word": "想", "start": 1414.267264423077, "end": 1414.427264423077, "probability": 0.99658203125 }, { "word": "了", "start": 1414.427264423077, "end": 1414.547264423077, "probability": 0.99169921875 }, { "word": "解", "start": 1414.547264423077, "end": 1414.687264423077, "probability": 1.0 }, { "word": "更", "start": 1414.687264423077, "end": 1414.8472644230771, "probability": 0.99853515625 }, { "word": "多", "start": 1414.8472644230771, "end": 1415.007264423077, "probability": 1.0 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 660, "review_annotation": "", "row_id": 660, "bounded_row_id": 660, "parent_row_ids": [ 660 ], "parent_row_id": 660, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "所以如果你们想了解更多", "selected_text": "所以如果你们想了解更多", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 660 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1415.007264423077, "end": 1416.227264423077, "text": "一定要去看看他们的论文", "raw_asr_text": "一定要去看看他们的论文", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 51, "words": [ { "word": "一定要", "start": 1415.007264423077, "end": 1415.247264423077, "probability": 0.95703125 }, { "word": "去", "start": 1415.247264423077, "end": 1415.447264423077, "probability": 0.994140625 }, { "word": "看看", "start": 1415.447264423077, "end": 1415.647264423077, "probability": 0.98095703125 }, { "word": "他们", "start": 1415.647264423077, "end": 1415.8272644230772, "probability": 0.94140625 }, { "word": "的", "start": 1415.8272644230772, "end": 1415.947264423077, "probability": 0.97509765625 }, { "word": "论", "start": 1415.947264423077, "end": 1416.0672644230772, "probability": 0.9990234375 }, { "word": "文", "start": 1416.0672644230772, "end": 1416.227264423077, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 661, "review_annotation": "", "row_id": 661, "bounded_row_id": 661, "parent_row_ids": [ 661 ], "parent_row_id": 661, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "一定要去看看他们的论文", "selected_text": "一定要去看看他们的论文", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 661 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1416.227264423077, "end": 1417.207264423077, "text": "谢谢大家观看", "raw_asr_text": "谢谢大家观看", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 51, "words": [ { "word": "谢谢", "start": 1416.227264423077, "end": 1416.547264423077, "probability": 0.951171875 }, { "word": "大家", "start": 1416.547264423077, "end": 1416.8272644230772, "probability": 0.99609375 }, { "word": "观", "start": 1416.8272644230772, "end": 1416.967264423077, "probability": 0.994140625 }, { "word": "看", "start": 1416.967264423077, "end": 1417.207264423077, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 662, "review_annotation": "", "row_id": 662, "bounded_row_id": 662, "parent_row_ids": [ 662 ], "parent_row_id": 662, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "谢谢大家观看", "selected_text": "谢谢大家观看", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 662 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1417.207264423077, "end": 1419.267264423077, "text": "特别感谢Spamatch Chris Ledoux", "raw_asr_text": "特别感谢Spamatch Chris Ledoux", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 51, "words": [ { "word": "特", "start": 1417.207264423077, "end": 1417.527264423077, "probability": 0.87939453125 }, { "word": "别", "start": 1417.527264423077, "end": 1417.6272644230771, "probability": 0.9990234375 }, { "word": "感", "start": 1417.6272644230771, "end": 1417.747264423077, "probability": 0.998046875 }, { "word": "谢", "start": 1417.747264423077, "end": 1417.927264423077, "probability": 0.9990234375 }, { "word": "Sp", "start": 1417.927264423077, "end": 1418.0672644230772, "probability": 0.456298828125 }, { "word": "am", "start": 1418.0672644230772, "end": 1418.207264423077, "probability": 0.58544921875 }, { "word": "atch", "start": 1418.207264423077, "end": 1418.427264423077, "probability": 0.343994140625 }, { "word": " Chris", "start": 1418.427264423077, "end": 1418.687264423077, "probability": 0.47900390625 }, { "word": " Led", "start": 1418.687264423077, "end": 1418.907264423077, "probability": 0.61767578125 }, { "word": "oux", "start": 1418.907264423077, "end": 1419.267264423077, "probability": 0.96142578125 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 663, "review_annotation": "", "row_id": 663, "bounded_row_id": 663, "parent_row_ids": [ 663 ], "parent_row_id": 663, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "特别感谢Spamatch Chris Ledoux", "selected_text": "特别感谢Spamatch Chris Ledoux", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 663 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1419.267264423077, "end": 1419.987264423077, "text": "Degan", "raw_asr_text": "Degan", "source_language": "en", "source_lang": "en", "asr_packet_id": 51, "words": [ { "word": "D", "start": 1419.267264423077, "end": 1419.6072644230771, "probability": 0.79443359375 }, { "word": "egan", "start": 1419.6072644230771, "end": 1419.987264423077, "probability": 0.8310546875 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 664, "review_annotation": "", "row_id": 664, "bounded_row_id": 664, "parent_row_ids": [ 664 ], "parent_row_id": 664, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "Degan", "selected_text": "Degan", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 664 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "en" }, { "start": 1419.987264423077, "end": 1421.8272644230772, "text": "Roberts Zaviasa", "raw_asr_text": "Roberts Zaviasa", "source_language": "en", "source_lang": "en", "asr_packet_id": 51, "words": [ { "word": "R", "start": 1419.987264423077, "end": 1420.487264423077, "probability": 0.8466796875 }, { "word": "ober", "start": 1420.487264423077, "end": 1420.6072644230771, "probability": 0.97412109375 }, { "word": "ts", "start": 1420.6072644230771, "end": 1420.8272644230772, "probability": 0.9990234375 }, { "word": " Z", "start": 1420.8272644230772, "end": 1421.007264423077, "probability": 0.41943359375 }, { "word": "av", "start": 1421.007264423077, "end": 1421.1272644230771, "probability": 0.7314453125 }, { "word": "ias", "start": 1421.1272644230771, "end": 1421.307264423077, "probability": 0.489990234375 }, { "word": "a", "start": 1421.307264423077, "end": 1421.8272644230772, "probability": 0.98779296875 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 665, "review_annotation": "", "row_id": 665, "bounded_row_id": 665, "parent_row_ids": [ 665 ], "parent_row_id": 665, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "Roberts Zaviasa", "selected_text": "Roberts Zaviasa", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 665 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "en" }, { "start": 1421.8272644230772, "end": 1422.447264423077, "text": "Poof", "raw_asr_text": "Poof", "source_language": "en", "source_lang": "en", "asr_packet_id": 51, "words": [ { "word": "P", "start": 1421.8272644230772, "end": 1422.1072644230771, "probability": 0.97216796875 }, { "word": "oof", "start": 1422.1072644230771, "end": 1422.447264423077, "probability": 0.8740234375 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 666, "review_annotation": "", "row_id": 666, "bounded_row_id": 666, "parent_row_ids": [ 666 ], "parent_row_id": 666, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "Poof", "selected_text": "Poof", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 666 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "en" }, { "start": 1422.447264423077, "end": 1423.167264423077, "text": "Inu", "raw_asr_text": "Inu", "source_language": "en", "source_lang": "en", "asr_packet_id": 51, "words": [ { "word": "In", "start": 1422.447264423077, "end": 1422.8472644230771, "probability": 0.2320556640625 }, { "word": "u", "start": 1422.8472644230771, "end": 1423.167264423077, "probability": 0.7060546875 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 667, "review_annotation": "", "row_id": 667, "bounded_row_id": 667, "parent_row_ids": [ 667 ], "parent_row_id": 667, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "Inu", "selected_text": "Inu", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 667 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "en" }, { "start": 1423.167264423077, "end": 1424.727264423077, "text": "DX Research Group", "raw_asr_text": "DX Research Group", "source_language": "en", "source_lang": "en", "asr_packet_id": 51, "words": [ { "word": "D", "start": 1423.167264423077, "end": 1423.6072644230771, "probability": 0.99072265625 }, { "word": "X", "start": 1423.6072644230771, "end": 1423.947264423077, "probability": 0.70068359375 }, { "word": " Research", "start": 1423.947264423077, "end": 1424.267264423077, "probability": 0.61083984375 }, { "word": " Group", "start": 1424.267264423077, "end": 1424.727264423077, "probability": 0.91845703125 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 668, "review_annotation": "", "row_id": 668, "bounded_row_id": 668, "parent_row_ids": [ 668 ], "parent_row_id": 668, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "DX Research Group", "selected_text": "DX Research Group", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 668 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "en" }, { "start": 1424.727264423077, "end": 1425.487264423077, "text": "Alex", "raw_asr_text": "Alex", "source_language": "en", "source_lang": "en", "asr_packet_id": 51, "words": [ { "word": "Alex", "start": 1424.727264423077, "end": 1425.487264423077, "probability": 0.97314453125 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 669, "review_annotation": "", "row_id": 669, "bounded_row_id": 669, "parent_row_ids": [ 669 ], "parent_row_id": 669, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "Alex", "selected_text": "Alex", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 669 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "en" }, { "start": 1425.487264423077, "end": 1426.8472644230771, "text": "Leo Halawani", "raw_asr_text": "Leo Halawani", "source_language": "en", "source_lang": "en", "asr_packet_id": 51, "words": [ { "word": "Le", "start": 1425.487264423077, "end": 1426.047264423077, "probability": 0.916015625 }, { "word": "o", "start": 1426.047264423077, "end": 1426.267264423077, "probability": 0.8681640625 }, { "word": " Hal", "start": 1426.267264423077, "end": 1426.407264423077, "probability": 0.7177734375 }, { "word": "aw", "start": 1426.407264423077, "end": 1426.547264423077, "probability": 0.703125 }, { "word": "ani", "start": 1426.547264423077, "end": 1426.8472644230771, "probability": 0.7294921875 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 670, "review_annotation": "", "row_id": 670, "bounded_row_id": 670, "parent_row_ids": [ 670 ], "parent_row_id": 670, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "Leo Halawani", "selected_text": "Leo Halawani", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 670 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "en" }, { "start": 1426.8472644230771, "end": 1428.807264423077, "text": "以及其他许多通过Patreon", "raw_asr_text": "以及其他许多通过Patreon", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 51, "words": [ { "word": "以及", "start": 1426.8472644230771, "end": 1427.167264423077, "probability": 0.97314453125 }, { "word": "其", "start": 1427.167264423077, "end": 1427.447264423077, "probability": 0.9921875 }, { "word": "他", "start": 1427.447264423077, "end": 1427.6072644230771, "probability": 0.95361328125 }, { "word": "许", "start": 1427.6072644230771, "end": 1427.767264423077, "probability": 0.9892578125 }, { "word": "多", "start": 1427.767264423077, "end": 1427.907264423077, "probability": 0.99951171875 }, { "word": "通", "start": 1427.907264423077, "end": 1428.047264423077, "probability": 0.9853515625 }, { "word": "过", "start": 1428.047264423077, "end": 1428.247264423077, "probability": 0.99853515625 }, { "word": "P", "start": 1428.247264423077, "end": 1428.407264423077, "probability": 0.86767578125 }, { "word": "atre", "start": 1428.407264423077, "end": 1428.547264423077, "probability": 0.986328125 }, { "word": "on", "start": 1428.547264423077, "end": 1428.807264423077, "probability": 0.99951171875 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 671, "review_annotation": "", "row_id": 671, "bounded_row_id": 671, "parent_row_ids": [ 671 ], "parent_row_id": 671, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "以及其他许多通过Patreon", "selected_text": "以及其他许多通过Patreon", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 671 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1428.807264423077, "end": 1430.967264423077, "text": "或YouTube支持我的人", "raw_asr_text": "或YouTube支持我的人", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 51, "words": [ { "word": "或", "start": 1428.807264423077, "end": 1429.307264423077, "probability": 0.92626953125 }, { "word": "You", "start": 1429.307264423077, "end": 1429.507264423077, "probability": 0.7919921875 }, { "word": "Tube", "start": 1429.507264423077, "end": 1429.807264423077, "probability": 0.99755859375 }, { "word": "支持", "start": 1429.807264423077, "end": 1430.3472644230771, "probability": 0.99169921875 }, { "word": "我的", "start": 1430.3472644230771, "end": 1430.687264423077, "probability": 0.99365234375 }, { "word": "人", "start": 1430.687264423077, "end": 1430.967264423077, "probability": 0.9951171875 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 672, "review_annotation": "", "row_id": 672, "bounded_row_id": 672, "parent_row_ids": [ 672 ], "parent_row_id": 672, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "或YouTube支持我的人", "selected_text": "或YouTube支持我的人", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 672 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1430.967264423077, "end": 1433.3272644230772, "text": "还没关注我的话到推特上关注我", "raw_asr_text": "还没关注我的话到推特上关注我", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 51, "words": [ { "word": "还", "start": 1430.967264423077, "end": 1431.3472644230771, "probability": 0.94287109375 }, { "word": "没", "start": 1431.3472644230771, "end": 1431.467264423077, "probability": 0.95947265625 }, { "word": "关", "start": 1431.467264423077, "end": 1431.6272644230771, "probability": 0.99951171875 }, { "word": "注", "start": 1431.6272644230771, "end": 1431.767264423077, "probability": 0.99755859375 }, { "word": "我的", "start": 1431.767264423077, "end": 1431.947264423077, "probability": 0.99169921875 }, { "word": "话", "start": 1431.947264423077, "end": 1432.1272644230771, "probability": 0.99951171875 }, { "word": "到", "start": 1432.1272644230771, "end": 1432.3272644230772, "probability": 0.97119140625 }, { "word": "推", "start": 1432.3272644230772, "end": 1432.447264423077, "probability": 0.95068359375 }, { "word": "特", "start": 1432.447264423077, "end": 1432.5672644230772, "probability": 0.99609375 }, { "word": "上", "start": 1432.5672644230772, "end": 1432.787264423077, "probability": 0.9970703125 }, { "word": "关", "start": 1432.787264423077, "end": 1432.967264423077, "probability": 0.986328125 }, { "word": "注", "start": 1432.967264423077, "end": 1433.147264423077, "probability": 0.99951171875 }, { "word": "我", "start": 1433.147264423077, "end": 1433.3272644230772, "probability": 0.9775390625 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 673, "review_annotation": "", "row_id": 673, "bounded_row_id": 673, "parent_row_ids": [ 673 ], "parent_row_id": 673, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "还没关注我的话到推特上关注我", "selected_text": "还没关注我的话到推特上关注我", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 673 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "zh" }, { "start": 1433.3272644230772, "end": 1434.447264423077, "text": "我们下期再见", "raw_asr_text": "我们下期再见", "source_language": "zh", "source_lang": "zh", "asr_packet_id": 51, "words": [ { "word": "我们", "start": 1433.3272644230772, "end": 1433.5872644230772, "probability": 0.96337890625 }, { "word": "下", "start": 1433.5872644230772, "end": 1433.8672644230771, "probability": 0.99658203125 }, { "word": "期", "start": 1433.8672644230771, "end": 1434.0672644230772, "probability": 0.9990234375 }, { "word": "再见", "start": 1434.0672644230772, "end": 1434.447264423077, "probability": 0.9990234375 } ], "asr_metrics": { "avg_logprob": -0.27720741892969886, "compression_ratio": 1.1564245810055866, "no_speech_prob": 4.596756308927752e-11 }, "lid_status": "STABLE", "alternatives": [ { "language": "zh", "text": "或者需要更多背景知识才能解释清楚比如DSEC缩放多莫态混合专家均衡以及其他基础设施改进所以如果你们想了解更多一定要去看看他们的论文谢谢大家观看特别感谢Spamatch Chris LedouxDeganRoberts ZaviasaPoofInuDX Research GroupAlexLeo Halawani以及其他许多通过Patreon或YouTube支持我的人还没关注我的话到推特上关注我我们下期再见" } ], "id": 674, "review_annotation": "", "row_id": 674, "bounded_row_id": 674, "parent_row_ids": [ 674 ], "parent_row_id": 674, "packet_id": 51, "LID_evidence": [ { "region_id": 0, "probabilities": { "zh": 0.9884244047511619 }, "status": "STABLE" } ], "VAD_region_ids": [ 0 ], "language_confidence": 0.9884244047511619, "ASR_engine": "mlx-whisper", "ASR_route": "bounded_vad_lid", "packet_start": 1406.187264423077, "packet_end": 1434.57525, "raw_text": "我们下期再见", "selected_text": "我们下期再见", "provenance_chain": [ { "stage": "bounded_asr", "parent_row_ids": [ 674 ], "packet_id": 51, "engine": "mlx-whisper" } ], "language": "zh" } ], "source": "production_bounded_vad_lid_asr", "audio_sha256": "c327f1e3f8a409fc393dc79c79e94284178df6af21a22dbeed7277a680e60e9f", "presegmentation_correction": { "applied": false, "correction_count": 0, "deterministic_fix_count": 0 } }