nemotron-gym-instruction-following-calendar
收藏资源简介:
本数据集是NVIDIA Nemotron-RL-Instruction-Following-Calendar-v2数据集的Harbor格式转换版本,属于NeMo-Gym集合的一部分。它专门设计用于强化学习任务,特别是指令跟随日历相关场景。数据集中每个样本包含两个核心字段:path字段为确定性短ID(格式为<family>-<sha256[:12]>.tar.gz),task_binary字段为包含完整Harbor任务的gzip压缩tar包。tar包内部遵循标准Harbor任务布局,包含关键文件如instruction.md(提示文本)、environment/Dockerfile(基础镜像和依赖)、tests/test.sh(验证器入口点)、tests/verifier.py(验证器实现)、tests/verifier_data.json(验证器输入数据)、metadata.json(数据溯源信息)和task.toml(任务配置)。数据集转换采用安全构建原则,确保内容不会插入到源代码中,所有值通过JSON文件在运行时解析。验证器家族为calendar_constraints,专门用于解析代理JSON列表并检查每个事件的持续时间、窗口和约束条件。数据集规模属于1K到10K类别,语言为英语,适用于强化学习研究和智能体训练。
This dataset is a Harbor format conversion of the NVIDIA Nemotron-RL-Instruction-Following-Calendar-v2 dataset, part of the NeMo-Gym collection. It is specifically designed for reinforcement learning tasks, particularly instruction-following calendar-related scenarios. Each sample in the dataset contains two core fields: the path field is a deterministic short ID (format <family>-<sha256[:12]>.tar.gz), and the task_binary field is a gzip-compressed tar package containing the complete Harbor task. Inside the tar package, it follows the standard Harbor task layout, including key files such as instruction.md (prompt text displayed to the agent), environment/Dockerfile (base image based on python:3.11-slim-bookworm and task-specific pip dependencies), tests/test.sh (validator entry point), tests/verifier.py (validator implementation), tests/verifier_data.json (validator input data for each task, in JSON format without code interpolation), metadata.json (data provenance information, including source dataset, row index, family, etc.), and task.toml (standard Harbor task configuration, including CPU/memory/timeout defaults). The dataset conversion process adopts secure construction principles, ensuring dataset content is not inserted into shell, Python, or Dockerfile source code, with all values parsed at runtime via JSON files. The validator family is calendar_constraints, specifically used to parse agent JSON lists and check the duration, window, and constraints of each event. The dataset size falls into the 1K<n<10K category, the language is English, and it is suitable for reinforcement learning research and agent training.




