nemotron-gym-multichallenge-advanced-v2
收藏资源简介:
该数据集名为laion/nemotron-gym-multichallenge-advanced-v2,是一个Harbor任务二进制数据集,专为文本生成任务设计,涉及智能体、港口操作、强化学习和Nemotron技术。它源自nvidia/Nemotron-RL-Multichallenge-v1 [advanced](属于nvidia/Nemotron-Post-Training-v3系列),包含1,068个任务。数据以两列形式组织:path(字符串类型)和task_binary(gzip压缩的tar文件),通过OpenThoughts-Agent的data.nemotron_gym框架转换而来。数据集主要用于智能体训练和评估,支持使用Rubric LLM judge(基于所有标准)进行自动化评分,试验过程中需要提供OPENAI_API_KEY。相较于前一版本,此版本修复了终端智能体的答案交付契约问题,通过明确指示智能体如何通过shell heredoc写入评分文件路径(并验证文件存在)来确保答案正确提交,而评分逻辑保持不变,从而提高了评估的准确性和可靠性。
This dataset, named `laion/nemotron-gym-multichallenge-advanced-v2`, is a Harbor task binary dataset specifically designed for text generation tasks, covering agents, harbor operations, reinforcement learning, and Nemotron technology. Derived from nvidia/Nemotron-RL-Multichallenge-v1 [advanced], which belongs to the nvidia/Nemotron-Post-Training-v3 series, it contains 1,068 tasks. The dataset is organized into two columns: `path` (string type) and `task_binary` (gzip-compressed tar file), and was converted via the `data.nemotron_gym` framework of OpenThoughts-Agent. It is primarily used for agent training and evaluation, supporting automated scoring with Rubric LLM judge based on all criteria. An `OPENAI_API_KEY` is required during the experimentation process. Compared to the previous version, this release fixes the answer delivery contract issue for terminal agents: it ensures correct answer submission by explicitly instructing agents on how to write the scoring file path via shell heredoc and verifying the file's existence, while keeping the scoring logic unchanged, thus improving the accuracy and reliability of evaluation.
数据集概述
- 数据集名称:
laion/nemotron-gym-multichallenge-advanced-v2 - 许可证: Apache-2.0
- 任务类别: 文本生成 (text-generation)
- 标签: agent, harbor, reinforcement-learning, nemotron
数据集内容
- 任务数量: 1,068 个任务
- 数据格式: 包含
path(字符串) 和task_binary(gzip tar 文件) 两个列 - 来源: 从
nvidia/Nemotron-RL-Multichallenge-v1 [advanced]转换而来,属于nvidia/Nemotron-Post-Training-v3集合的一部分 - 转换框架: 使用 OpenThoughts-Agent
data.nemotron_gym框架进行转换
评分机制
- 评分方式: 使用 Rubric LLM 评分器(评估所有标准)
- 运行要求: 运行试验时需要设置
OPENAI_API_KEY
版本变更说明(相比前一版本)
- 核心修复: 修复了终端智能体的答案交付协议问题
- 问题描述: 前一版本仅指示智能体生成答案,但未指定如何提交;导致
terminus-2智能体在单轮交互中通过聊天回复输出答案,而未将结果写入评分文件,使得大多数试验因“答案文件缺失”而得到 0 分 - 改进方案: 新指令明确要求智能体通过 shell heredoc 将答案写入评分器的文件路径(并验证写入成功),评分逻辑本身未作其他更改




