nemotron-gym-math-v4
收藏资源简介:
laion/nemotron-gym-math-v4 是一个用于智能体任务的数据集,源自 nvidia/Nemotron-RL-Math-v2 并整合了 DAPO-Math-17k 和 Skywork-OR1 指针,属于 nvidia/Nemotron-Post-Training-v3 集合的一部分。该数据集包含 6,534 个任务,数据格式包括两个列:path(字符串类型)和 task_binary(gzip tar 格式),通过 OpenThoughts-Agent 的 data.nemotron_gym 框架转换。数据集适用于文本生成和强化学习任务,特别是与智能体相关的数学问题解决或模拟环境。评估方法基于 sympy/latex oxed{} 的等价性对比黄金标准进行评分。与前一版本相比,此版本修复了终端智能体的答案提交契约,明确指示通过 shell heredoc 将答案写入评分文件路径并验证,以解决答案文件缺失导致的评分问题。数据集标签包括 agent、harbor、reinforcement-learning 和 nemotron,采用 Apache 2.0 许可证。
laion/nemotron-gym-math-v4 is a dataset for agent tasks, derived from nvidia/Nemotron-RL-Math-v2 and integrated with DAPO-Math-17k and Skywork-OR1 pointers, forming part of the nvidia/Nemotron-Post-Training-v3 collection. It contains 6,534 tasks, with data format including two columns: path (string type) and task_binary (gzip tar format), converted via the OpenThoughts-Agents data.nemotron_gym framework. The dataset is suitable for text generation and reinforcement learning tasks, particularly for agent-related mathematical problem solving or simulation environments. Evaluation is based on scoring through equivalence comparison of sympy/latex oxed{} against a gold standard. Compared to the previous version, this release fixes the answer submission contract for terminal agents, explicitly instructing to write answers to the scoring file path via shell heredoc and validate, to address scoring issues caused by missing answer files. Dataset labels include agent, harbor, reinforcement-learning, and nemotron, under the Apache 2.0 license.
数据集概述
数据集名称:laion/nemotron-gym-math-v4
许可证:Apache-2.0
任务类别:文本生成(text-generation)
标签:agent, harbor, reinforcement-learning, nemotron
规模:6,534 个任务(Harbor task-binary 格式)
来源与构建:从 nvidia/Nemotron-RL-Math-v2 转换而来,并合并了 hydrated DAPO-Math-17k / Skywork-OR1 pointers(属于 nvidia/Nemotron-Post-Training-v3 系列)。转换过程使用 OpenThoughts-Agent 的 data.nemotron_gym 框架,数据包含两列:path(字符串类型)和 task_binary(gzip tar 格式)。
评分方式:基于 sympy/latex 的 oxed{} 等价性评估,与标准答案(gold)进行比较。
版本变更说明(v4 相比前序版本)
- 核心改进:修复了终端智能体的 答案提交协议。
- 问题背景:前序版本仅告诉智能体 输出什么内容,但未明确 如何提交。导致 1 回合的
terminus-2智能体将答案以聊天回复的形式输出,而非写入评分文件,多数试次因“答案文件缺失”得分为 0。 - 改进措施:新版本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,并执行验证,确保答案文件正确写入。评分逻辑本身未作改变。




