laion/nemotron-gym-knowledge-mcqa

该数据集是nvidia/Nemotron-RL-knowledge-mcqa的Harbor格式转换版本，属于NVIDIA的NeMo-Gym集合的一部分。它主要用于强化学习任务，涉及知识多选问答（MCQA）。数据集包含100K到1M条记录，每行数据包括path（确定性短ID）和task_binary（包含完整Harbor任务的gzip压缩tar文件）。tarball内容遵循Harbor任务布局，包括指令文件、环境Dockerfile、验证器脚本、验证器数据JSON文件、元数据JSON文件和任务配置TOML文件。转换过程安全，确保数据不插入到shell、Python或Dockerfile源中，所有值通过JSON文件传递，基础镜像固定，文本字段经过清理，tarball路径验证安全，且tarball是确定性的。数据集可用于Harbor平台运行任务，支持通过Python脚本提取和运行。

This dataset is the Harbor-format converted version of nvidia/Nemotron-RL-knowledge-mcqa, and is part of NVIDIA's NeMo-Gym collection. It is primarily intended for reinforcement learning tasks, focusing on multiple-choice knowledge question answering (MCQA). The dataset contains between 100,000 and 1,000,000 records. Each row includes two fields: "path" (a deterministic short ID) and "task_binary" (a gzip-compressed tar archive containing a complete Harbor task). The contents of the tarball adhere to the standard Harbor task layout, including instruction files, environment Dockerfiles, validator scripts, validator data JSON files, metadata JSON files, and task configuration TOML files. The conversion process is secure: no data is injected into shell, Python, or Dockerfile sources, all values are passed via JSON files, the base image is fixed, text fields are sanitized, tarball paths are verified for safety, and the tarball is deterministic. This dataset can be used to run tasks on the Harbor platform, and supports extraction and execution via Python scripts.