nemotron-gym-agentic-function-calling-pivot-v2
收藏资源简介:
该数据集是一个Harbor任务二进制数据集,包含9,579个任务,由nvidia/Nemotron-RL-Agentic-Function-Calling-Pivot-v1转换而来,属于nvidia/Nemotron-Post-Training-v3集合的一部分。数据集使用OpenThoughts-Agent的data.nemotron_gym框架进行转换,主要包含两列数据:path(字符串类型)和task_binary(gzip tar格式)。该数据集设计用于单步工具调用匹配(function_call)或LLM判断(message)的评估任务。与前一版本相比,v2版本修复了终端代理的答案交付契约问题:明确指示代理通过shell heredoc将答案写入评分器指定的文件路径并进行验证,而不仅仅是告知需要生产什么内容,解决了之前版本因代理以聊天回复形式输出答案导致评分文件缺失的问题。数据集适用于文本生成任务,特别关注代理、强化学习等应用场景。
This dataset is a Harbor task binary dataset containing 9,579 tasks, converted from nvidia/Nemotron-RL-Agentic-Function-Calling-Pivot-v1, and is part of the nvidia/Nemotron-Post-Training-v3 collection. It was converted using the data.nemotron_gym framework from OpenThoughts-Agent, and primarily includes two columns: `path` (string type) and `task_binary` (gzip tar format). This dataset is designed for evaluation tasks involving single-step tool call matching (function_call) or LLM judgment (message). Compared to its prior version, the v2 release fixes the answer delivery contract issue for terminal agents: it explicitly instructs agents to write answers to the file path specified by the scorer via shell heredoc and conduct verification, rather than merely notifying what content needs to be generated. This resolves the problem of missing scoring files in previous versions caused by agents outputting answers in the form of chat replies. The dataset is applicable to text generation tasks, with a particular focus on application scenarios such as AI agents and reinforcement learning.
数据集概述
数据集名称:laion/nemotron-gym-agentic-function-calling-pivot-v2
许可证:Apache-2.0
任务类别:文本生成
标签:agent, harbor, reinforcement-learning, nemotron
数据集规模与来源
- 任务数量:9,579个任务(Harbor任务二进制格式)
- 来源:基于 nvidia/Nemotron-RL-Agentic-Function-Calling-Pivot-v1 转换而来,属于 nvidia/Nemotron-Post-Training-v3 的一部分
数据格式
- 列结构:包含
path(字符串)和task_binary(gzip tar格式) - 转换框架:使用 OpenThoughts-Agent 的
data.nemotron_gym框架完成转换
评估方式(Grading)
- 单步评估:基于工具调用匹配(function_call)或LLM裁判(message)进行评分
版本变更说明
- 改进目标:修复终端代理的答案交付合约问题
- 先前版本问题:旧版本仅告知代理交付什么,但未说明如何提交;导致1回合的
terminus-2代理以聊天回复形式输出答案,而非写入评分文件,因此多数试炼因“答案文件缺失”得分为0 - 本版本改进:指令明确要求通过 shell heredoc 将答案写入评分文件路径(并进行验证),评分逻辑无其他变化




