{"text": "大家好,我是大叔大,先算笔账,让大模型做判断,它得一个字一个字把答案写出来,几百毫秒就没了,你还得再写正则去解析它写的东西。更要命的是,它哪天换个说法,你整条流程当场崩。今天要聊的Laya正好,反过来33毫秒,一次前向传播,结构化,答案已经递到你手上了。这期我不讲虚的,就带你从零把它跑起来。被环境跑第一个决策制建服务,一步步来,Laya是个非自回归的决策模型。思路其实很简单,把需求写成决策,一次前向传播直接给结构化答案它在开源社区已经攒了18000多个star,单次决策33毫秒,支持100多种语言,用的是Apache 2.0协议。先搞清楚它和生成式模型差在哪,这点想明白了。后面全通生成式的做法,你肯定熟,提示词发过去,等它生成一段文本再写正则或者Jason解析,最后还得防它答非所问。Mans一方面,不稳才是大问题。Laya完全不同,它把状态和问题一起喂进去,一次前向传播,直接拿到Choice Score Now三种结构化答案,它一共就三个决策原语。Choice是选择从若干选项里挑一个适合部门路由,分类,主题归类。Score是打分,在有序档位上给个分值,适合紧急度,挫败感,危害程度。No是是非判断,给一个是或否的,校准概率,适合钓鱼,骚扰信息,流失风险。说白了,分类用Choice,程度用Score。真伟用No一次钓用,几个问题一起答,知道它是干什么的了,第一步先把本地环境备好。这块门槛真的低,只要Python 3.10或者更高版本,依赖就是Hugging Face,客户端,Transformers和Pytorch这一套,公开模型第一次拉取,不用账号,具体三步,先建钢3.11,按MV.win,然后用环境里的Python执行钢MP,installLayer,把Layer装上,最后执行Python钢I钢C导入Layer并答应版本号,有三点要注意,一是用独立的虚拟环境把依赖圈在一起,别污染全局的Python,二是装和跑,必须用同一个解释器,不然会爆找不到Layer,三是第一次会从Hugging Face拉模型,之后就走本地缓存,对了,参数里的钢I会把当前目录从导入境里排除,这样本地随便放的一份代码不会假装你已经装好了,环境通过,通了第二步,三分钟跑通第一个决策,推荐从旅游模式入手,我们拿一封中文客服邮件当例子,一次调用,回答三个问题,先fromLayerImportRouter,然后Router等于Router,括号里Preload等于处,预加载Checkpoint,让首次旅游就能秒回,接着写State,也就是要分析的那段文本,我这个月的会员被扣了两次钱,请把多扣的马上退给我,不然我就退定了,然后写Questions,用中文描述你要问的决策,一个问题叫部门类型是Choice,问的是这条请求,该交给哪个部门处理,Criteria里列了四个候选不错,销售对应报价,新合同,其他就是以上都不是,另一个叫紧急度,类型是Score,问这条请求有多紧急,档位从低到高是不及,尽快处理,已经主射业务,还有一个叫流失风险,类型是No,问用户是不是威胁要退订或者离开,最后Rest等于Router点,Predict,把State和Questions传进去,一次调用,拿到全部答案,打印出来看看,部门是账单,紧急度1.84,满分2.0,流失风险0.89,整套动作就三件事,见Router,写State,写Questions,问题选项,说明全用中文�顺带把三种原语的问法讲透,先看Answers,每个问题一条答案还带着概率和自信度,比如部门这条Choice是账单,Confidence是0.94,Probabilities里把每个选项的概率都列了出来,账单0.94,技术0.03,销售0.02,其他0.01,再看Routing,这是路由原数据,告诉你这次实际用了哪个Checkpoint,以及为什么Model是实际选中的Checkpoint,Repo是对应的权中目录,Reason是选它的理由,也就是脚本和语言判断的结果,三种原语怎么问呢,Choice就给一组带说明的候选,Criteria里每一项写清含义,让他选,Score就给一条从低到高的单位列表返回期望单位,No只问一件事成不成立,返回0到1之间的校准概率,这里我得单独提一句Confidence真不是摆设,它是用严格适当评分规则训练出来的,可以直接拿来当门控,后面细讲,接下来是个实际问题,到底用Router还是直接加载模型,Router会在推理之前,先用不到0.5毫秒判断一件事,英文Checkpoint能不能读懂这段话,为什么非要提前判断,因为英文模型碰上非拉丁字母会自信的答错,它不会提醒所以这个判断只能放在推理前,路由模式的常见写法是这样,Router等于Router,Prehead等于True,默认常驻两个checkpoint,非英文流量为主的话,可以设Default等于Multilingual,改掉默认语言,你已经有语言识别模型了,直接用Langloose等于这把结果交给他,想省内存设Maxload等于1,代价是每次切换都要再加载一遍,不用了,就条Router点Unload释放内存,只服务一种语言的话,单模型直连更省心,Agent等于Layer点Load,传ConveyInnovations,斜杠Layer,这是英文版,加载多语言版加载结构化决策工作流,就换成Subfolder等于Typed-Decisions,三个checkpoint怎么挑,Layer用的是ModernBird-Large,4亿2100万参数,上下文512,适合英文场景,Layer-Multilingual用的是HalmyBearER-Base,3亿2200万参数,上下文1024,支持100多种语言,速度快一倍,Layer-Typed-Decisions,也是ModernBird-Large,4亿2100万参数,上下文1024,适合结构化决策工作流,不想从零设计问题的话,官方还预调好了,4套schema导入就能用,另外也能把整套能力挂成HTPP接口,4套预设分别是这样,模型路由用Layer.Router-Chack-Stations,判断这条请求该给小模型还是旗舰模型,比如用Ely,注入重构这个服务,就该交给更强的模型,提示词护栏,用Layer.Guard-Chack-Stations,拦住月域,注入和泄露,比如提示词里写忽略之前所有指令,内容安全,用Layer.Moderation-Chack-Stations,识别读信骚扰,威胁类内容,工单分整,用Layer.Tree-Edge-Chack-Stations,一次拿Layer.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.AA.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.AA.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.AA.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 30.0, "text": "大家好,我是大叔大,先算笔账,让大模型做判断,它得一个字一个字把答案写出来,几百毫秒就没了,你还得再写正则去解析它写的东西。更要命的是,它哪天换个说法,你整条流程当场崩。今天要聊的Laya正好,反过来33毫秒,一次前向传播,结构化,答案已经递到你手上了。这期我不讲虚的,就带你从零把它跑起来。被环境跑第一个决策制建服务,一步步来,Laya是个非自回归的决策模型。思路其实很简单,把需求写成决策,一次前向传播直接给结构化答案", "tokens": [50365, 15248, 11, 15914, 3582, 39936, 3582, 11, 10108, 19497, 5437, 242, 18464, 99, 11, 33650, 3582, 41908, 39823, 10907, 2437, 97, 4307, 255, 11, 11284, 5916, 20182, 22381, 20182, 22381, 16075, 28223, 28899, 5676, 247, 44561, 11, 6336, 254, 31906, 7256, 104, 20956, 3111, 10062, 2289, 11, 2166, 14852, 5916, 8623, 5676, 247, 15789, 2437, 247, 6734, 17278, 7360, 238, 11284, 5676, 247, 1546, 38409, 16220, 1543, 19002, 4275, 25236, 24620, 11, 11284, 17028, 6135, 26075, 95, 7549, 8090, 11148, 11, 2166, 27662, 48837, 27854, 29649, 16233, 50255, 49378, 102, 1543, 12074, 4275, 40096, 1546, 43, 4427, 15789, 2131, 11, 22138, 16866, 6912, 10191, 7256, 104, 20956, 11, 27505, 8945, 24282, 7384, 254, 49993, 11, 45641, 7360, 226, 23756, 11, 28223, 28899, 49161, 2215, 240, 4511, 2166, 11389, 5708, 2289, 1543, 5562, 16786, 25395, 39255, 12026, 248, 1546, 11, 3111, 4845, 99, 2166, 35630, 6306, 114, 42061, 32585, 9147, 6912, 1543, 23238, 8051, 107, 47885, 32585, 18049, 7549, 5676, 111, 7973, 244, 25491, 34157, 27408, 5087, 94, 11, 2257, 31429, 31429, 6912, 11, 43, 4427, 1541, 7549, 12107, 9722, 8350, 7391, 240, 1546, 5676, 111, 7973, 244, 41908, 39823, 1543, 8870, 24658, 41646, 4563, 11249, 222, 47446, 11, 16075, 32535, 32718, 5676, 247, 11336, 5676, 111, 7973, 244, 11, 27505, 8945, 24282, 7384, 254, 49993, 43297, 23197, 45641, 7360, 226, 23756, 28223, 28899], "temperature": 0, "avg_logprob": -0.057919883728027345, "compression_ratio": 1.3951807228915662, "no_speech_prob": 8.206062565574435e-12}, {"id": 1, "seek": 3000, "start": 30.0, "end": 60.0, "text": "它在开源社区已经攒了18000多个star,单次决策33毫秒,支持100多种语言,用的是Apache 2.0协议。先搞清楚它和生成式模型差在哪,这点想明白了。后面全通生成式的做法,你肯定熟,提示词发过去,等它生成一段文本再写正则或者Jason解析,最后还得防它答非所问。Mans一方面,不稳才是大问题。Laya完全不同,它把状态和问题一起喂进去,一次前向传播,直接拿到Choice Score Now三种结构化答案,它一共就三个决策原语。Choice是选择从若干选项里挑一个适合部门路由,", "tokens": [50365, 11284, 3581, 18937, 47402, 27658, 9937, 118, 49161, 5917, 240, 2289, 6494, 1360, 6392, 7549, 9710, 11, 47446, 9487, 5676, 111, 7973, 244, 10191, 7256, 104, 20956, 11, 35488, 6879, 6392, 39810, 5233, 255, 12009, 11, 9254, 24620, 32, 79, 6000, 568, 13, 15, 5322, 237, 7422, 106, 1543, 10108, 41816, 38433, 11284, 12565, 8244, 11336, 27584, 41908, 39823, 21679, 3581, 17028, 11, 5562, 12579, 7093, 28430, 2289, 1543, 13547, 8833, 11319, 19550, 8244, 11336, 27584, 1546, 10907, 11148, 11, 2166, 14356, 107, 12088, 21805, 253, 11, 20949, 25696, 5233, 235, 28926, 16866, 6734, 11, 10187, 11284, 8244, 11336, 2257, 28427, 17174, 8802, 8623, 5676, 247, 15789, 2437, 247, 31148, 38267, 17278, 7360, 238, 11, 8661, 13547, 14852, 5916, 35863, 11284, 28223, 12107, 5966, 22064, 1543, 44, 599, 2257, 39604, 11, 1960, 10415, 111, 18888, 1541, 3582, 34069, 1543, 43, 4427, 37100, 47123, 11, 11284, 16075, 35276, 114, 3757, 223, 12565, 34069, 29567, 35142, 36700, 6734, 11, 27505, 8945, 24282, 7384, 254, 49993, 11, 43297, 24351, 4511, 44473, 573, 47901, 823, 10960, 39810, 45641, 7360, 226, 23756, 28223, 28899, 11, 11284, 2257, 15408, 3111, 10960, 7549, 5676, 111, 7973, 244, 19683, 5233, 255, 1543, 44473, 573, 1541, 2215, 231, 6852, 102, 35630, 13736, 98, 26111, 2215, 231, 10178, 117, 15759, 49067, 20182, 2215, 224, 14245, 13470, 8259, 101, 24658, 23786, 11], "temperature": 0, "avg_logprob": -0.06988818335426228, "compression_ratio": 1.2259414225941423, "no_speech_prob": 1.1669040733686131e-11}, {"id": 2, "seek": 6000, "start": 60.0, "end": 90.0, "text": "分类,主题归类。Score是打分,在有序档位上给个分值,适合紧急度,挫败感,危害程度。No是是非判断,给一个是或否的,校准概率,适合钓鱼,骚扰信息,流失风险。说白了,分类用Choice,程度用Score。真伟用No一次钓用,几个问题一起答,知道它是干什么的了,第一步先把本地环境备好。这块门槛真的低,只要Python 3.10或者更高版本,依赖就是Hugging Face,客户端,Transformers和Pytorch这一套,公开模型第一次拉取,不用账号,具体三步,先建", "tokens": [50365, 6627, 22113, 119, 11, 13557, 30716, 7391, 240, 22113, 119, 1543, 50, 12352, 1541, 12467, 6627, 11, 3581, 2412, 6346, 237, 21416, 96, 11160, 5708, 23197, 7549, 6627, 40242, 11, 2215, 224, 14245, 7732, 100, 36379, 13127, 11, 8501, 104, 18464, 98, 9709, 11, 38292, 14694, 29649, 13127, 1543, 4540, 1541, 1541, 12107, 2437, 97, 4307, 255, 11, 23197, 20182, 1541, 19780, 45334, 1546, 11, 49684, 6336, 228, 26181, 44866, 11, 2215, 224, 14245, 25153, 241, 165, 109, 120, 11, 49657, 248, 3416, 108, 17665, 26460, 11, 27854, 28564, 47209, 8842, 102, 1543, 8090, 13558, 2289, 11, 6627, 22113, 119, 9254, 44473, 573, 11, 29649, 13127, 9254, 50, 12352, 1543, 6303, 7384, 253, 9254, 4540, 27505, 25153, 241, 9254, 11, 6336, 254, 7549, 34069, 29567, 28223, 11, 7758, 11284, 1541, 26111, 10440, 1546, 2289, 11, 18049, 31429, 10108, 16075, 8802, 10928, 8051, 107, 47885, 1787, 229, 2131, 1543, 5562, 47734, 8259, 101, 24880, 249, 8034, 41377, 11, 14003, 4275, 47, 88, 11943, 805, 13, 3279, 31148, 19002, 12979, 42096, 8802, 11, 3254, 251, 5266, 244, 5620, 39, 697, 3249, 4047, 11, 32316, 1486, 115, 11957, 107, 11, 33339, 837, 433, 12565, 47, 4328, 284, 339, 5562, 2257, 1881, 245, 11, 13545, 18937, 41908, 39823, 18049, 9487, 29717, 29436, 11, 24384, 18464, 99, 26987, 11, 39806, 29485, 10960, 31429, 11, 10108, 34157], "temperature": 0, "avg_logprob": -0.0670175552368164, "compression_ratio": 1.2102908277404922, "no_speech_prob": 1.3750049709937429e-11}, {"id": 3, "seek": 9000, "start": 90.0, "end": 120.0, "text": "钢3.11,按MV.win,然后用环境里的Python执行钢MP,installLayer,把Layer装上,最后执行Python钢I钢C导入Layer并答应版本号,有三点要注意,一是用独立的虚拟环境把依赖圈在一起,别污染全局的Python,二是装和跑,必须用同一个解释器,不然会爆找不到Layer,三是第一次会从Hugging Face拉模型,之后就走本地缓存,对了,参数里的钢I会把当前目录从导入境里排除,这样本地随便放的一份代码不会假装你已经装好了,环境通过,", "tokens": [50365, 25153, 95, 18, 13, 5348, 11, 26613, 44, 53, 13, 9136, 11, 26636, 9254, 8051, 107, 47885, 15759, 1546, 47, 88, 11943, 3416, 100, 8082, 25153, 95, 12224, 11, 40166, 43, 11167, 11, 16075, 43, 11167, 8083, 227, 5708, 11, 8661, 13547, 3416, 100, 8082, 47, 88, 11943, 25153, 95, 40, 25153, 95, 34, 4510, 120, 14028, 43, 11167, 3509, 114, 28223, 44297, 42096, 8802, 26987, 11, 2412, 10960, 12579, 4275, 40794, 11, 2257, 1541, 9254, 18637, 105, 24409, 1546, 12026, 248, 6852, 253, 8051, 107, 47885, 16075, 3254, 251, 5266, 244, 2523, 230, 3581, 29567, 11, 18453, 12800, 94, 46568, 11319, 34703, 1546, 47, 88, 11943, 11, 11217, 1541, 8083, 227, 12565, 32585, 11, 28531, 10178, 119, 9254, 13089, 20182, 17278, 5873, 232, 34386, 11, 1960, 5823, 12949, 45827, 25085, 35728, 43, 11167, 11, 10960, 1541, 18049, 9487, 12949, 35630, 39, 697, 3249, 4047, 29717, 41908, 39823, 11, 9574, 13547, 3111, 9575, 8802, 10928, 38109, 241, 39781, 11, 8713, 2289, 11, 2129, 224, 33188, 15759, 1546, 25153, 95, 40, 12949, 16075, 16233, 8945, 11386, 7391, 243, 35630, 4510, 120, 14028, 47885, 15759, 44647, 32999, 11, 21209, 8802, 10928, 10673, 237, 27364, 12744, 1546, 2257, 36266, 19105, 23230, 223, 47928, 31706, 8083, 227, 2166, 49161, 8083, 227, 12621, 11, 8051, 107, 47885, 19550, 16866, 11], "temperature": 0, "avg_logprob": -0.0984791259062455, "compression_ratio": 1.2446043165467626, "no_speech_prob": 1.5100280401081356e-11}, {"id": 4, "seek": 12000, "start": 120.0, "end": 150.0, "text": "通了第二步,三分钟跑通第一个决策,推荐从旅游模式入手,我们拿一封中文客服邮件当例子,一次调用,回答三个问题,先fromLayerImportRouter,然后Router等于Router,括号里Preload等于处,预加载Checkpoint,让首次旅游就能秒回,接着写State,也就是要分析的那段文本,我这个月的会员被扣了两次钱,请把多扣的马上退给我,不然我就退定了,然后写Questions,用中文描述你要问的决策,一个问题叫部门类型是Choice,问的是这条请求,该交给哪个部门处理,Criteria里列了四个候选", "tokens": [50365, 19550, 2289, 19693, 31429, 11, 10960, 6627, 50064, 32585, 19550, 18049, 7549, 5676, 111, 7973, 244, 11, 33597, 31409, 238, 35630, 4479, 227, 9592, 116, 41908, 27584, 14028, 11389, 11, 15003, 24351, 2257, 1530, 223, 5975, 17174, 32316, 27408, 3023, 106, 20485, 16233, 17797, 7626, 11, 27505, 8897, 225, 9254, 11, 8350, 28223, 10960, 7549, 34069, 11, 10108, 20579, 43, 11167, 43329, 477, 49, 23985, 11, 26636, 49, 23985, 10187, 37732, 49, 23985, 11, 40322, 26987, 15759, 42844, 2907, 10187, 37732, 1787, 226, 11, 12501, 226, 9990, 17819, 121, 47769, 6053, 11, 33650, 25444, 9487, 4479, 227, 9592, 116, 3111, 8225, 20956, 8350, 11, 14468, 20708, 5676, 247, 4520, 473, 11, 6404, 5620, 4275, 6627, 7360, 238, 1546, 4184, 28427, 17174, 8802, 11, 1654, 15368, 6939, 1546, 12949, 3606, 246, 23238, 3416, 96, 2289, 36257, 9487, 39623, 11, 27908, 16075, 6392, 3416, 96, 1546, 49970, 5708, 46361, 49076, 11, 1960, 5823, 22020, 46361, 12088, 2289, 11, 26636, 5676, 247, 8547, 377, 626, 11, 9254, 5975, 17174, 11673, 237, 3316, 108, 32085, 22064, 1546, 5676, 111, 7973, 244, 11, 20182, 34069, 19855, 13470, 8259, 101, 22113, 119, 39823, 1541, 44473, 573, 11, 22064, 24620, 5562, 48837, 27908, 32718, 11, 44646, 28455, 23197, 17028, 7549, 13470, 8259, 101, 1787, 226, 13876, 11, 38750, 1681, 654, 15759, 43338, 2289, 19425, 7549, 11238, 2215, 231], "temperature": 0, "avg_logprob": -0.046026118667670965, "compression_ratio": 1.2662337662337662, "no_speech_prob": 1.2607940733100342e-11}, {"id": 5, "seek": 15000, "start": 150.0, "end": 180.0, "text": "不错,销售对应报价,新合同,其他就是以上都不是,另一个叫紧急度,类型是Score,问这条请求有多紧急,档位从低到高是不及,尽快处理,已经主射业务,还有一个叫流失风险,类型是No,问用户是不是威胁要退订或者离开,最后Rest等于Router点,Predict,把State和Questions传进去,一次调用,拿到全部答案,打印出来看看,部门是账单,紧急度1.84,满分2.0,流失风险0.89,整套动作就三件事,见Router,写State,写Questions,问题选项,说明全用中文�", "tokens": [50365, 1960, 29900, 11, 23049, 222, 17198, 106, 8713, 44297, 49817, 1550, 115, 11, 12560, 14245, 13089, 11, 9572, 5000, 5620, 29497, 7182, 7296, 11, 22762, 20182, 19855, 7732, 100, 36379, 13127, 11, 22113, 119, 39823, 1541, 50, 12352, 11, 22064, 5562, 48837, 27908, 32718, 2412, 6392, 7732, 100, 36379, 11, 21416, 96, 11160, 35630, 41377, 4511, 12979, 1541, 1960, 25703, 11, 1530, 121, 10251, 1787, 226, 13876, 11, 49161, 13557, 1530, 226, 940, 248, 5087, 94, 11, 35091, 20182, 19855, 27854, 28564, 47209, 8842, 102, 11, 22113, 119, 39823, 1541, 4540, 11, 22064, 9254, 1486, 115, 23034, 19976, 223, 164, 17486, 4275, 46361, 7422, 95, 31148, 18614, 119, 18937, 11, 8661, 13547, 49, 377, 10187, 37732, 49, 23985, 12579, 11, 47, 24945, 11, 16075, 4520, 473, 12565, 8547, 377, 626, 7384, 254, 36700, 6734, 11, 27505, 8897, 225, 9254, 11, 24351, 4511, 38714, 28223, 28899, 11, 12467, 35825, 44561, 22884, 11, 13470, 8259, 101, 1541, 18464, 99, 47446, 11, 7732, 100, 36379, 13127, 16, 13, 25494, 11, 15868, 94, 6627, 17, 13, 15, 11, 27854, 28564, 47209, 8842, 102, 15, 13, 21115, 11, 27662, 1881, 245, 34961, 11914, 3111, 10960, 20485, 6973, 11, 23813, 49, 23985, 11, 5676, 247, 4520, 473, 11, 5676, 247, 8547, 377, 626, 11, 34069, 2215, 231, 10178, 117, 11, 8090, 11100, 11319, 9254, 5975, 17174, 5676], "temperature": 0, "avg_logprob": -0.06670742291506096, "compression_ratio": 1.2511627906976743, "no_speech_prob": 1.4521579251580707e-11}, {"id": 6, "seek": 18000, "start": 180.0, "end": 209.34, "text": "顺带把三种原语的问法讲透,先看Answers,每个问题一条答案还带着概率和自信度,比如部门这条Choice是账单,Confidence是0.94,Probabilities里把每个选项的概率都列了出来,账单0.94,技术0.03,销售0.02,其他0.01,再看Routing,这是路由原数据,告诉你这次实际用了哪个Checkpoint,以及为什么Model是实际选中的Checkpoint,Repo是对应的权中目录,Reason是选它的理由,也就是脚本和语言判断的结果,三种原语怎么问呢,Choice就给一组带说明的候选,", "tokens": [50365, 10178, 118, 4845, 99, 16075, 10960, 39810, 19683, 5233, 255, 1546, 22064, 11148, 39255, 45808, 11, 10108, 4200, 32, 3695, 5364, 11, 23664, 7549, 34069, 2257, 48837, 28223, 28899, 14852, 4845, 99, 20708, 26181, 44866, 12565, 9722, 17665, 13127, 11, 36757, 13470, 8259, 101, 5562, 48837, 44473, 573, 1541, 18464, 99, 47446, 11, 43874, 2778, 1541, 15, 13, 27032, 11, 12681, 65, 6167, 15759, 16075, 23664, 7549, 2215, 231, 10178, 117, 1546, 26181, 44866, 7182, 43338, 2289, 44561, 11, 18464, 99, 47446, 15, 13, 27032, 11, 32502, 1474, 107, 15, 13, 11592, 11, 23049, 222, 17198, 106, 15, 13, 12756, 11, 9572, 5000, 15, 13, 10607, 11, 8623, 4200, 49, 24500, 11, 27455, 24658, 23786, 19683, 33188, 26075, 106, 11, 44816, 2166, 5562, 9487, 24726, 8842, 227, 9254, 2289, 17028, 7549, 47769, 6053, 11, 40282, 31006, 44, 41147, 1541, 24726, 8842, 227, 2215, 231, 5975, 1546, 47769, 6053, 11, 25554, 78, 1541, 8713, 44297, 1546, 4422, 225, 5975, 11386, 7391, 243, 11, 8524, 1258, 1541, 2215, 231, 45224, 13876, 23786, 11, 6404, 5620, 27067, 248, 8802, 12565, 5233, 255, 12009, 2437, 97, 4307, 255, 1546, 45641, 9319, 11, 10960, 39810, 19683, 5233, 255, 15282, 22064, 6240, 11, 44473, 573, 3111, 23197, 2257, 10115, 226, 4845, 99, 8090, 11100, 1546, 11238, 2215, 231, 11, 51832], "temperature": 0, "avg_logprob": -0.054584903101767265, "compression_ratio": 1.2616822429906542, "no_speech_prob": 1.515268466256714e-11}, {"id": 7, "seek": 20934, "start": 209.34, "end": 239.34, "text": "Criteria里每一项写清含义,让他选,Score就给一条从低到高的单位列表返回期望单位,No只问一件事成不成立,返回0到1之间的校准概率,这里我得单独提一句Confidence真不是摆设,它是用严格适当评分规则训练出来的,可以直接拿来当门控,后面细讲,接下来是个实际问题,到底用Router还是直接加载模型,Router会在推理之前,先用不到0.5毫秒判断一件事,英文Checkpoint能不能读懂这段话,为什么非要提前判断,因为英文模型碰上非拉丁字母会自信的答错,它不会提醒", "tokens": [50365, 38750, 1681, 654, 15759, 23664, 2257, 10178, 117, 5676, 247, 21784, 2392, 104, 2930, 231, 11, 33650, 5000, 2215, 231, 11, 50, 12352, 3111, 23197, 2257, 48837, 35630, 41377, 4511, 12979, 1546, 47446, 11160, 43338, 17571, 3316, 242, 8350, 16786, 22694, 47446, 11160, 11, 4540, 14003, 22064, 2257, 20485, 6973, 11336, 1960, 11336, 24409, 11, 3316, 242, 8350, 15, 4511, 16, 9574, 31685, 1546, 49684, 6336, 228, 26181, 44866, 11, 35102, 1654, 5916, 47446, 18637, 105, 20949, 2257, 34592, 43874, 2778, 6303, 7296, 34783, 228, 7422, 122, 11, 11284, 1541, 9254, 940, 98, 30921, 2215, 224, 16233, 5233, 226, 6627, 6758, 226, 2437, 247, 7422, 255, 10115, 225, 44561, 1546, 11, 6723, 43297, 24351, 6912, 16233, 8259, 101, 48707, 11, 13547, 8833, 10115, 228, 39255, 11, 14468, 4438, 6912, 1541, 7549, 24726, 8842, 227, 34069, 11, 33883, 9254, 49, 23985, 45726, 43297, 9990, 17819, 121, 41908, 39823, 11, 49, 23985, 12949, 3581, 33597, 13876, 32442, 11, 10108, 9254, 35728, 15, 13, 20, 7256, 104, 20956, 2437, 97, 4307, 255, 2257, 20485, 6973, 11, 27869, 17174, 47769, 6053, 8225, 28590, 5233, 119, 29624, 5562, 28427, 21596, 11, 31006, 12107, 4275, 20949, 8945, 2437, 97, 4307, 255, 11, 34627, 27869, 17174, 41908, 39823, 16337, 108, 5708, 12107, 29717, 940, 223, 22381, 35744, 12949, 9722, 17665, 1546, 28223, 29900, 11, 11284, 47928, 20949, 40796], "temperature": 0, "avg_logprob": -0.03985388717309242, "compression_ratio": 1.2723311546840959, "no_speech_prob": 1.3964840568647663e-11}, {"id": 8, "seek": 23934, "start": 239.34, "end": 269.34000000000003, "text": "所以这个判断只能放在推理前,路由模式的常见写法是这样,Router等于Router,Prehead等于True,默认常驻两个checkpoint,非英文流量为主的话,可以设Default等于Multilingual,改掉默认语言,你已经有语言识别模型了,直接用Langloose等于这把结果交给他,想省内存设Maxload等于1,代价是每次切换都要再加载一遍,不用了,就条Router点Unload释放内存,只服务一种语言的话,单模型直连更省心,Agent等于Layer点Load,传ConveyInnovations,斜杠Layer,这是英文版,加载多语言版", "tokens": [50365, 7239, 15368, 2437, 97, 4307, 255, 14003, 8225, 12744, 3581, 33597, 13876, 8945, 11, 24658, 23786, 41908, 27584, 1546, 11279, 23813, 5676, 247, 11148, 1541, 21209, 11, 49, 23985, 10187, 37732, 49, 23985, 11, 42844, 1934, 10187, 37732, 14252, 622, 11, 6173, 246, 7422, 97, 11279, 24023, 119, 36257, 7549, 15723, 6053, 11, 12107, 27869, 17174, 27854, 26748, 13992, 13557, 44575, 11, 6723, 7422, 122, 35, 5666, 5107, 10187, 37732, 44, 723, 38219, 11, 34490, 29327, 6173, 246, 7422, 97, 5233, 255, 12009, 11, 2166, 49161, 2412, 5233, 255, 12009, 5233, 228, 18453, 41908, 39823, 2289, 11, 43297, 9254, 43, 656, 752, 541, 10187, 37732, 5562, 16075, 45641, 9319, 28455, 23197, 5000, 11, 7093, 2862, 223, 34742, 39781, 7422, 122, 36025, 2907, 10187, 37732, 16, 11, 19105, 1550, 115, 1541, 23664, 9487, 23632, 26075, 95, 7182, 4275, 8623, 9990, 17819, 121, 2257, 3330, 235, 11, 24384, 2289, 11, 3111, 48837, 49, 23985, 12579, 12405, 2907, 5873, 232, 12744, 34742, 39781, 11, 14003, 27408, 5087, 94, 2257, 39810, 5233, 255, 12009, 44575, 11, 47446, 41908, 39823, 16186, 3316, 252, 19002, 2862, 223, 7945, 11, 32, 6930, 10187, 37732, 43, 11167, 12579, 31645, 345, 11, 7384, 254, 9838, 5603, 4575, 4185, 763, 11, 4307, 250, 4422, 254, 43, 11167, 11, 27455, 27869, 17174, 42096, 11, 9990, 17819, 121, 6392, 5233, 255, 12009, 42096], "temperature": 0, "avg_logprob": -0.0693588214070273, "compression_ratio": 1.2776523702031604, "no_speech_prob": 1.4692663752313706e-11}, {"id": 9, "seek": 26934, "start": 269.34, "end": 299.34, "text": "加载结构化决策工作流,就换成Subfolder等于Typed-Decisions,三个checkpoint怎么挑,Layer用的是ModernBird-Large,4亿2100万参数,上下文512,适合英文场景,Layer-Multilingual用的是HalmyBearER-Base,3亿2200万参数,上下文1024,支持100多种语言,速度快一倍,Layer-Typed-Decisions,也是ModernBird-Large,4亿2100万参数,上下文1024,适合结构化决策工作流,不想从零设计问题的话,官方还预调好了,", "tokens": [50365, 9990, 17819, 121, 45641, 7360, 226, 23756, 5676, 111, 7973, 244, 41315, 27854, 11, 3111, 26075, 95, 11336, 39582, 18353, 260, 10187, 37732, 22464, 3452, 12, 35, 3045, 4252, 11, 10960, 7549, 15723, 6053, 15282, 49067, 11, 43, 11167, 9254, 24620, 44, 378, 1248, 33, 1271, 12, 43, 289, 432, 11, 19, 1369, 123, 4436, 628, 23570, 2129, 224, 33188, 11, 5708, 4438, 17174, 20, 4762, 11, 2215, 224, 14245, 27869, 17174, 50255, 50218, 11, 43, 11167, 12, 44, 723, 38219, 9254, 24620, 39, 304, 2226, 6524, 289, 1598, 12, 33, 651, 11, 18, 1369, 123, 7490, 628, 23570, 2129, 224, 33188, 11, 5708, 4438, 17174, 3279, 7911, 11, 35488, 6879, 6392, 39810, 5233, 255, 12009, 11, 31217, 13127, 10251, 2257, 35477, 11, 43, 11167, 12, 22464, 3452, 12, 35, 3045, 4252, 11, 22021, 44, 378, 1248, 33, 1271, 12, 43, 289, 432, 11, 19, 1369, 123, 4436, 628, 23570, 2129, 224, 33188, 11, 5708, 4438, 17174, 3279, 7911, 11, 2215, 224, 14245, 45641, 7360, 226, 23756, 5676, 111, 7973, 244, 41315, 27854, 11, 1960, 7093, 35630, 6306, 114, 7422, 122, 7422, 94, 34069, 44575, 11, 31929, 9249, 14852, 12501, 226, 8897, 225, 12621, 11], "temperature": 0, "avg_logprob": -0.07663481368631275, "compression_ratio": 1.3768996960486322, "no_speech_prob": 1.619917741613186e-11}, {"id": 10, "seek": 29934, "start": 299.34, "end": 329.34, "text": "4套schema导入就能用,另外也能把整套能力挂成HTPP接口,4套预设分别是这样,模型路由用Layer.Router-Chack-Stations,判断这条请求该给小模型还是旗舰模型,比如用Ely,注入重构这个服务,就该交给更强的模型,提示词护栏,用Layer.Guard-Chack-Stations,拦住月域,注入和泄露,比如提示词里写忽略之前所有指令,内容安全,用Layer.Moderation-Chack-Stations,识别读信骚扰,威胁类内容,工单分整,用Layer.Tree-Edge-Chack-Stations,一次拿", "tokens": [50365, 19, 1881, 245, 6145, 5619, 4510, 120, 14028, 3111, 8225, 9254, 11, 26202, 6404, 8225, 16075, 27662, 1881, 245, 8225, 13486, 8501, 224, 11336, 39, 51, 17755, 14468, 18144, 11, 19, 1881, 245, 12501, 226, 7422, 122, 6627, 18453, 1541, 21209, 11, 41908, 39823, 24658, 23786, 9254, 43, 11167, 13, 49, 23985, 12, 6546, 501, 12, 4520, 763, 11, 2437, 97, 4307, 255, 5562, 48837, 27908, 32718, 44646, 23197, 7322, 41908, 39823, 45726, 4479, 245, 10256, 108, 41908, 39823, 11, 36757, 9254, 36, 356, 11, 26432, 14028, 12624, 7360, 226, 15368, 27408, 5087, 94, 11, 3111, 44646, 28455, 23197, 19002, 5702, 118, 1546, 41908, 39823, 11, 20949, 25696, 5233, 235, 7235, 97, 8987, 237, 11, 9254, 43, 11167, 13, 38, 16981, 12, 6546, 501, 12, 4520, 763, 11, 6852, 99, 21632, 6939, 16262, 253, 11, 26432, 14028, 12565, 6847, 226, 18594, 110, 11, 36757, 20949, 25696, 5233, 235, 15759, 5676, 247, 4263, 121, 6904, 98, 32442, 39300, 25922, 49061, 11, 34742, 25750, 45466, 11, 9254, 43, 11167, 13, 44, 378, 5053, 12, 6546, 501, 12, 4520, 763, 11, 5233, 228, 18453, 5233, 119, 17665, 49657, 248, 3416, 108, 11, 19976, 223, 164, 17486, 22113, 119, 34742, 25750, 11, 23323, 47446, 6627, 27662, 11, 9254, 43, 11167, 13, 51, 701, 12, 27061, 432, 12, 6546, 501, 12, 4520, 763, 11, 27505, 24351], "temperature": 0, "avg_logprob": -0.0952563521038791, "compression_ratio": 1.3534031413612566, "no_speech_prob": 1.6977351816049868e-11}, {"id": 11, "seek": 32934, "start": 329.34, "end": 359.34, "text": "Layer.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.", "tokens": [50365, 43, 11167, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13], "temperature": 0, "avg_logprob": -0.05420141690515082, "compression_ratio": 13.176470588235293, "no_speech_prob": 1.7168939880907175e-11}, {"id": 12, "seek": 35934, "start": 359.34, "end": 389.34, "text": "A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A", "tokens": [50365, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32], "temperature": 0, "avg_logprob": -0.01862762006408965, "compression_ratio": 17.0, "no_speech_prob": 1.0456420451721016e-10}, {"id": 13, "seek": 38934, "start": 389.34, "end": 419.34, "text": "A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A", "tokens": [50365, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32], "temperature": 0, "avg_logprob": -0.01688198765297107, "compression_ratio": 17.0, "no_speech_prob": 1.1933770638350438e-10}, {"id": 14, "seek": 41934, "start": 419.34, "end": 449.34, "text": "A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A", "tokens": [50365, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32], "temperature": 0, "avg_logprob": -0.018840896709082906, "compression_ratio": 17.0, "no_speech_prob": 1.0167584829634535e-10}, {"id": 15, "seek": 44934, "start": 449.34, "end": 470.94, "text": "A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A.A", "tokens": [50365, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32, 13, 32], "temperature": 0, "avg_logprob": -0.02115341152310906, "compression_ratio": 17.0, "no_speech_prob": 1.6167082950158118e-10}], "language": "zh"}