laion/nemotron-gym-knowledge-openqa
收藏Hugging Face2026-05-16 更新2026-05-31 收录
下载链接:
https://hf-mirror.com/datasets/laion/nemotron-gym-knowledge-openqa
下载链接
链接失效反馈官方服务:
资源简介:
该数据集是nvidia/Nemotron-RL-knowledge-openqa的Harbor格式转换版本,用于强化学习任务。每行数据包含path(确定性短ID)和task_binary(包含完整Harbor任务的gzip压缩tar文件)。Harbor任务包括instruction.md(显示给智能体的提示)、environment/Dockerfile(基于python:3.11-slim-bookworm的Docker镜像)、tests/test.sh(验证器入口点)、tests/verifier.py(验证器实现)、tests/verifier_data.json(每个任务的验证器输入JSON文件)、metadata.json(来源信息)和task.toml(标准Harbor任务配置)。转换过程安全构造,确保数据集内容不插入到shell、Python或Dockerfile源代码中,所有值通过JSON文件传递,且tarball路径经过验证以防止攻击。数据集源自NVIDIA的NeMo-Gym集合,适用于Harbor环境中的知识开放问答任务。
Harbor-format conversion of nvidia/Nemotron-RL-knowledge-openqa. Each row contains: path (deterministic short ID) and task_binary (gzipped tar containing the full Harbor task). The tarball contents follow Harbors task layout: instruction.md (prompt shown to the agent), environment/Dockerfile (python:3.11-slim-bookworm base + task-specific pip dependencies), tests/test.sh (verifier entrypoint), tests/verifier.py (verifier implementation), tests/verifier_data.json (per-task verifier inputs in JSON), metadata.json (provenance information), and task.toml (standard Harbor task config). Conversion is secure-by-construction: dataset content is never interpolated into shell, Python, or Dockerfile source, all values flow through JSON, and tarball paths are validated against attacks. Derived from NVIDIAs NeMo-Gym collection, for knowledge open QA tasks in reinforcement learning.
提供机构:
laion


