laion/nemotron-gym-knowledge-openqa-v2

该数据集是nvidia/Nemotron-RL-knowledge-openqa的Harbor格式转换版本，用于强化学习中的知识开放问答任务。每行数据包含path（确定性短ID）和task_binary（gzip压缩的tar文件，内含完整的Harbor任务），任务文件包括instruction.md（代理提示）、环境配置、验证器实现和元数据等。转换过程安全可靠，确保数据不注入代码，使用llm_judge验证器（基于LiteLLM，默认openai/gpt-4o-mini）根据参考答案进行语义等价评分。数据集源自NVIDIA的NeMo-Gym集合，适用于Harbor平台运行强化学习任务。

This dataset is a Harbor-format conversion of nvidia/Nemotron-RL-knowledge-openqa, designed for reinforcement learning tasks in knowledge open question answering. Each row contains a path (deterministic short ID) and task_binary (a gzipped tar file with the full Harbor task), which includes files such as instruction.md (prompt for the agent), environment configuration, verifier implementation, and metadata. The conversion is secure-by-construction, preventing code interpolation, and uses an llm_judge verifier (based on LiteLLM, defaulting to openai/gpt-4o-mini) for semantic-equivalence scoring against reference answers. Derived from NVIDIAs NeMo-Gym collection, it is suitable for running reinforcement learning tasks on the Harbor platform.