HKBU-KnowComp/patchworld-trajectories
收藏资源简介:
该数据集名为PatchWorld Trajectory Splits,是用于复现PatchWorld论文实验(RQ1/RQ2离线归纳和评估)的轨迹数据。它包含七个AgentGym环境(alfworld、babyai、maze、sciworld、textcraft、webshop、wordle),每个环境都提供了训练、验证和测试的JSONL格式分割文件。每个轨迹数据包括元数据(如环境、项目ID、任务索引、成功状态等)和转换列表(观察、动作、下一个观察、奖励、完成状态)。数据集旨在支持离线强化学习研究,特别是世界模型和智能体轨迹分析。
The dataset is named PatchWorld Trajectory Splits and contains trajectory data for reproducing experiments from the PatchWorld paper (RQ1/RQ2 offline induction and evaluation). It includes seven AgentGym environments (alfworld, babyai, maze, sciworld, textcraft, webshop, wordle), each with train, validation, and test splits in JSONL format. Each trajectory consists of metadata (e.g., env, item_id, task_idx, success) and a list of transitions (observation, action, next_observation, reward, done). The dataset is designed to support offline reinforcement learning research, particularly for world models and agent trajectory analysis.




