。然后“We present Navital Gym...”这部分是介绍Navital Gym,然后讲任务的评分规则,然后模型测试结果,然后额外的内容(私有集、7B adapter、许可、规则说明)。不对,要保留所有信息,不能增删。 首先,先处理: 第一个自然段落:AI agents are being deployed...another language model. 用
包裹,其中重要的比如“Public agent evaluations rarely cover these workflows, and many score with another language model.”?不对,重要的句子可以加strong?比如Navital Gym是重点,所以We present Navital Gym, a set of six open environments with 321 public tasks in five task families: trip planning, UPI payments and disputes, GST invoicing and returns, form filling (including a multi-document split) and customer support in Hindi, Hinglish and three experimental Indian languages.?不对,要自然。 然后,接下来的“Every task is graded by named, deterministic rules that recompute the outcome from a synthetic sandbox world. No language model judges any score.”可以用
,其中“No language model judges any score.”可以加strong? 然后模型测试部分:“We report one run per task for three models on all six environments, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 4.5, and for Gemini 3.1 Flash-Lite on two (UPI and the Support core). Opus solves 315 of 321 tasks, Sonnet 302 and Haiku 205. Trip planning and Hindi/Hinglish support separate the models most.”用
,结果部分可以加strong?比如Opus solves 315 of 321 tasks, Sonnet 302 and Haiku 205.还有Trip planning and Hindi/Hinglish support separate the models most. 然后剩下的部分:“We also describe a private held-out set that is not published, and a 7B adapter trained in the gym. Code is Apache-2.0 and task data is CC BY 4.0. The rules are a documented simplification of Indian practice, not the live systems.”用
?不对,有没有需要列表的?哦,任务族的部分,原文里“five task families: trip planning, UPI payments and disputes, GST invoicing and returns, form filling (including a multi-document split) and customer support in Hindi, Hinglish and three experimental Indian languages.”这里可以用
- 来列任务族?对哦,规则里说列表内容用ul li,那这部分可以处理成列表,这样更语义化。 哦对,刚才没注意,任务族是列表内容,所以可以调整: 首先,
AI agents are being deployed on Indian workflows such as payments, tax, forms, travel and customer support in Hindi and Hinglish. Public agent evaluations rarely cover these workflows, and many score with another language model.
然后We present Navital Gym, a set of six open environments with 321 public tasks in five task families:
然后- trip planning
- UPI payments and disputes
- GST invoicing and returns
- form filling (including a multi-document split)
- customer support in Hindi, Hinglish and three experimental Indian languages
Every task is graded by named, deterministic rules that recompute the outcome from a synthetic sandbox world. No language model judges any score.
然后We report one run per task for three models on all six environments, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 4.5, and for Gemini 3.1 Flash-Lite on two (UPI and the Support core). Opus solves 315 of 321 tasks, Sonnet 302 and Haiku 205. Trip planning and Hindi/Hinglish support separate the models most.
然后We also describe a private held-out set that is not published, and a 7B adapter trained in the gym. Code is Apache-2.0 and task data is CC BY 4.0. The rules are a documented simplification of Indian practice, not the live systems.
不对,有没有需要h2/h3的?原文是一篇新闻,有没有小节?比如可以分几个h2?比如“背景与现有评估缺口”、“Navital Gym框架”、“任务评分规则”、“模型测试结果”、“补充说明”?不对,不行,不能自己加内容,h2的文本必须是原文里的?哦不对,规则说“保留原文所有信息,不得增删改内容”,所以h2的标题必须是原文里的内容,不能自己编。哦对哦,不能自己加标题,因为会增内容。那原文里没有现成的小标题,那是不是不用h2/h3?不对,再看原文,有没有可以划分的小节?比如原文的结构是:1. 背景(AI agents部署,评估问题);2. 提出Navital Gym,任务族;3. 评分规则;4. 模型测试结果;5. 补充内容。但没有现成的标题,所以不能自己加h2的文本,因为会增内容。那可以不用h2/h3吗?规则说“长文用/
划分小节”,如果原文没有现成的小标题,就不需要硬加,因为不能增改内容。 哦对,绝对不能加原文没有的内容,所以h2/h3只有原文有对应的标题才用,这里原文没有,所以不用。那刚才的处理: 首先第一段:
AI agents are being deployed on Indian workflows such as payments, tax, forms, travel and customer support in Hindi and Hinglish. Public agent evaluations rarely cover these workflows, and many score with another language model.
然后介绍Navital Gym,其中任务族用列表:We present Navital Gym, a set of six open environments with 321 public tasks in five task families:
- trip planning
- UPI payments and disputes
- GST invoicing and returns
- form filling (including a multi-document split)
- customer support in Hindi, Hinglish and three experimental Indian languages
Every task is graded by named, deterministic rules that recompute the outcome from a synthetic sandbox world. No language model judges any score.
然后模型测试部分:We report one run per task for three models on all six environments, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 4.5, and for Gemini 3.1 Flash-Lite on two (UPI and the Support core). Opus solves 315 of 321 tasks, Sonnet 302 and Haiku 205. Trip planning and Hindi/Hinglish support separate the models most.
然后补充部分:We also describe a private held-out set that is not published, and a 7B adapter trained in the gym. Code is Apache-2.0 and task data is CC BY 4.0. The rules are a documented simplification of Indian practice, not the live systems.
等下,检查有没有增删改内容:原文的“five task families:”后面的内容拆成li,没有改,只是列表化,符合规则。重要的句子加了strong,比如Navital Gym的介绍,评分的关键句,测试结果,都是合理的,没有改内容。有没有需要引用或者代码的?没有,原文没有引用和代码