patchworld-trajectories
收藏资源简介:
PatchWorld Trajectory Splits 是一个用于强化学习研究的轨迹数据集,旨在支持论文《PatchWorld: Gradient-Free Optimization of Executable World Models》中RQ1和RQ2的离线世界模型归纳与评估任务的复现。该数据集包含来自七个不同AgentGym环境(alfworld、babyai、maze、sciworld、textcraft、webshop、wordle)的交互轨迹。每个环境均提供了标准的训练集、验证集和测试集划分,以JSONL格式存储。每个轨迹样本包含详细的元数据(如环境标识、任务索引、回合索引、数据划分、成功标志、总奖励)以及一系列状态转换记录(包括观察、动作、下一个观察、奖励和完成标志)。数据集规模在1万到10万条轨迹之间,适用于代理行为分析、世界模型学习、离线强化学习等研究场景。
PatchWorld Trajectory Splits is a trajectory dataset for reinforcement learning research, designed to support the reproduction of offline world model induction and evaluation tasks in RQ1 and RQ2 of the paper PatchWorld: Gradient-Free Optimization of Executable World Models. The dataset contains interaction trajectories from seven different AgentGym environments (alfworld, babyai, maze, sciworld, textcraft, webshop, wordle). Each environment provides standard train, validation, and test splits, stored in JSONL format. Each trajectory sample includes detailed metadata (such as environment identifier, task index, episode index, data split, success flag, total reward) and a series of state transition records (including observation, action, next observation, reward, and done flag). The dataset size ranges from 10,000 to 100,000 trajectories, suitable for research scenarios like agent behavior analysis, world model learning, and offline reinforcement learning.
数据集概述
数据集名称:PatchWorld Trajectory Splits
发布机构:HKBU-KnowComp
用途:用于复现论文《PatchWorld: Gradient-Free Optimization of Executable World Models》(RQ1/RQ2中的离线归纳与评估)的轨迹数据。
数据集规模与配置
- 语言:英文
- 许可协议:MIT
- 样本数量:10,000 < N < 100,000
- 任务类型:强化学习
- 标签:agent、world-model、trajectories、agentgym、patchworld
环境与文件结构
数据集包含 7个AgentGym环境,每个环境提供训练/验证/测试三个JSONL文件:
| 环境目录 | 说明 |
|---|---|
alfworld/ |
AlfWorld环境 |
babyai/ |
BabyAI环境 |
maze/ |
迷宫环境 |
sciworld/ |
SciWorld环境 |
textcraft/ |
TextCraft环境 |
webshop/ |
WebShop环境 |
wordle/ |
Wordle环境 |
每个目录下包含:
<env>_traj_train.jsonl<env>_traj_val.jsonl<env>_traj_test.jsonl
此外,manifest.json 文件列出了各文件的校验和与轨迹数量。
数据格式
每行一条轨迹,采用JSONL格式,每条轨迹包含:
metadata:包括env(环境名称)、item_id、task_idx、rollout_index、split(数据集划分)、success(是否成功)、total_reward(总奖励)transitions:一个列表,每个元素包含observation(观测)、action(动作)、next_observation(下一观测)、reward(奖励)、done(是否结束)
下载方式
方式一:从PatchWorld仓库根目录运行: bash bash scripts/download_data.sh
方式二:使用Hugging Face CLI直接下载: bash hf download HKBU-KnowComp/patchworld-trajectories --repo-type dataset --local-dir artifacts/patchworld/data_release --exclude "*.tar.gz"
使用示例
下载后,在仓库根目录可运行如下命令(以maze环境为例)进行模型归纳与评估:
bash patchworld-benchmark --env maze --train_glob "artifacts/patchworld/data_release/maze/maze_traj_train.jsonl" --eval_glob "artifacts/patchworld/data_release/maze/maze_traj_test.jsonl" --output_dir artifacts/patchworld/results
详细指令请参考 PatchWorld README 与 EXPERIMENTS.md。
引用信息
bibtex @misc{bai2026patchworldgradientfreeoptimizationexecutable, title={PatchWorld: Gradient-Free Optimization of Executable World Models}, author={Jiaxin Bai and Yue Guo and Yifei Dong and Jiaxuan Xiong and Tianshi Zheng and Yixia Li and Tianqing Fang and Yufei Li and Yisen Gao and Haoyu Huang and Zhongwei Xie and Hong Ting Tsang and Zihao Wang and Lihui Liu and Jeff Pan and Yangqiu Song}, year={2026}, eprint={2605.30880}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2605.30880}, }




