nemotron-gym-instruction-following
收藏资源简介:
本数据集是nvidia/Nemotron-RL-instruction_following数据集的Harbor格式转换版本,属于NVIDIA NeMo-Gym集合的一部分,专门用于强化学习任务,特别是指令遵循场景。数据集规模在1万到10万样本之间,语言为英语。每个数据样本包含两个字段:path(确定性短ID字符串)和task_binary(包含完整Harbor任务的gzip压缩tar二进制数据)。Harbor任务遵循标准布局,包括向智能体展示的提示指令文件(instruction.md)、基于python:3.11-slim-bookworm并添加任务特定依赖的环境Dockerfile、验证器入口点脚本(test.sh)及其实现(verifier.py)、每任务的验证器输入数据(verifier_data.json)、记录来源的元数据文件(metadata.json)以及标准的Harbor任务配置文件(task.toml)。转换过程注重安全性,确保数据集内容不会插值到shell、Python或Dockerfile源代码中,所有值通过JSON文件传递;基础镜像固定;文本字段经过控制字符清理和长度限制;tarball路径经过防遍历等攻击验证;且tarball具有确定性以保证可复现性。验证器家族采用IFEval风格的约束检查器,检查段落、单词、禁止内容和格式等。数据集适用于基于Harbor平台的强化学习训练和评估,用户可以通过Hugging Face datasets库加载数据,并提取单个任务在Harbor环境中运行。
This dataset is a Harbor format conversion of the nvidia/Nemotron-RL-instruction_following dataset, part of the NVIDIA NeMo-Gym collection, specifically designed for reinforcement learning tasks, particularly instruction-following scenarios. It contains between 10,000 and 100,000 samples in English. Each sample includes two fields: path (a deterministic short ID string) and task_binary (gzip-compressed tar binary data containing a complete Harbor task). Harbor tasks follow a standard layout, including a prompt instruction file (instruction.md) shown to the agent, an environment Dockerfile based on python:3.11-slim-bookworm with task-specific dependencies, a verifier entry point script (test.sh) and its implementation (verifier.py), per-task verifier input data (verifier_data.json), a metadata file (metadata.json) recording the source, and a standard Harbor task configuration file (task.toml). The conversion emphasizes security, ensuring dataset content is not interpolated into shell, Python, or Dockerfile source code, with all values passed via JSON files; the base image is fixed; text fields are cleaned of control characters and length-limited; tarball paths are validated against traversal and other attacks; and tarballs are deterministic for reproducibility. The verifier family uses IFEval-style constraint checkers to inspect paragraphs, words, prohibited content, and formatting. The dataset is suitable for reinforcement learning training and evaluation on the Harbor platform, and users can load the data via the Hugging Face datasets library to extract and run individual tasks in a Harbor environment.




