plfb
收藏资源简介:
本数据集为Policy Learning from Tutorial Books via Understanding, Rehearsing and Introspecting (URI)研究提供配套生成产物,旨在支持相关代码发布的复现工作。数据集主要包含两部分核心内容:一是针对Google Research Football(足球)环境的离线强化学习轨迹数据、由URI方法生成的想象演练数据以及大语言模型(LLM)的处理结果;二是针对Tic-Tac-Toe(井字棋)游戏的教程与模拟器相关产物。此外,数据集还提供了精选的最终模型检查点、评估日志以及详细的文件清单。数据总规模为29,668个文件,约37.03 GB。具体目录包括足球离线数据集、足球想象数据集、足球LLM生成结果、井字棋数据以及各类评估和模型文件。数据集适用于离线强化学习、基于教程的策略学习等研究任务。请注意,原始RedPajama书籍文本未包含在此公开版本中,且数据集内涉及的第三方模拟器代码和资产需遵循其原有许可。
This dataset is the supporting generated data artifact for the research titled *Policy Learning from Tutorial Books via Understanding, Rehearsing and Introspecting (URI)*, aiming to support the reproducibility of the associated code release. The dataset mainly consists of two core parts: 1) Offline reinforcement learning trajectory data for the Google Research Football (GRF) environment, imagined rehearsal data generated by the URI method, and processing results from Large Language Models (LLMs); 2) Tutorials and simulator-related artifacts for the Tic-Tac-Toe game. In addition, the dataset also provides curated final model checkpoints, evaluation logs, and a detailed file inventory. The dataset includes a total of 29,668 files and has an overall size of approximately 37.03 GB. The specific directory structure encompasses the football offline dataset, football imagined rehearsal dataset, LLM-generated results for the football environment, Tic-Tac-Toe data, and various evaluation and model files. This dataset is applicable to research tasks including offline reinforcement learning and tutorial-based policy learning. Please note that the original RedPajama book texts are not included in this public release, and any third-party simulator code and assets included in the dataset must adhere to their original licenses.
数据集概述:Policy Learning from Tutorial Books Data (PLFB)
该数据集是 Policy Learning from Tutorial Books via Understanding, Rehearsing and Introspecting (URI) 论文的配套数据,用于复现公开的 PLfB/URI 代码。数据集中包含 Football 和井字棋 (Tic-Tac-Toe) 两大领域的生成产物与训练评估工件。
数据集信息
- 许可证:Apache-2.0
- 任务类别:强化学习 (Reinforcement Learning)
- 语言:英文
- 标签:强化学习、离线强化学习、语言模型、Google Research Football、井字棋
数据集内容
数据集共包含 29,668 个文件,总大小约 37.03 GB,按目录分类如下:
| 目录 | 文件数 | 大小 | 内容说明 |
|---|---|---|---|
football/offline_dataset-v4/ |
26,421 | 36.06 GB | Football 离线轨迹、BC/ORL 缓冲区、代码/语言标签 |
football/imaginary_dataset_0204/ |
66 | 630.00 MB | URI 生成的 Football 排练数据 |
football/generated_llm_results/ |
3,062 | 200.90 MB | 生成的 LLM 输出(排除调试视频) |
tictactoe/ |
49 | 61.07 MB | 井字棋教程/模拟器工件 |
artifacts/football/final_uri_best/ |
18 | 47.42 MB | 最佳 Football URI 检查点、参数和评估 CSV |
artifacts/football/paper_uri_seed60/ |
51 | 61.18 KB | 论文 URI seed60 评估 CSV 和参数 |
artifacts/tictactoe/final_model/ |
1 | 28.73 MB | 井字棋最终模型检查点 |
目录结构
football/offline_dataset-v4/ football/offline_dataset-v4/sample_data/ football/imaginary_dataset_0204/ football/generated_llm_results/ tictactoe/ artifacts/football/final_uri_best/ artifacts/football/paper_uri_seed60/ artifacts/tictactoe/final_model/ .gitattributes manifests/inventory.jsonl manifests/inventory_summary.json
加载注意事项
.npz文件可能包含对象数组,需使用numpy.load(..., allow_pickle=True)加载。- 部分
.pkl文件为 HDF5 容器,需使用h5py打开。 - 部分
.jsonl文件为多行 JSON 字典格式,非严格的 JSON Lines。 - 可使用验证脚本
validate_data_release.py检查数据集的完整性。
数据路径与环境变量
通过 inspect_plfb_dataset.py 脚本可导出以下环境变量:
PLFB_DATASET_PATHPLFB_OFFLINE_DATASET_PATHPLFB_SAMPLED_DATA_PATHPLFB_IMAGINARY_DATASET_PATHPLFB_GENERATED_LLM_RESULTS_PATHPLFB_TTT_DATA_PATH
也可直接设置 PLFB_HF_DATASET_ROOT 指向下载的数据集根目录。
引用
bibtex @inproceedings{NEURIPS2024_21cf8411, author = {Chen, Xiong-Hui and Wang, Ziyan and Du, Yali and Jiang, Shengyi and Fang, Meng and Yu, Yang and Wang, Jun}, booktitle = {Advances in Neural Information Processing Systems}, doi = {10.52202/079017-0600}, editor = {A. Globerson and L. Mackey and D. Belgrave and A. Fan and U. Paquet and J. Tomczak and C. Zhang}, pages = {18940--18987}, publisher = {Curran Associates, Inc.}, title = {Policy Learning from Tutorial Books via Understanding, Rehearsing and Introspecting}, url = {https://proceedings.neurips.cc/paper_files/paper/2024/file/21cf8411ed825614e00006a1d9aab7e4-Paper-Conference.pdf}, volume = {37}, year = {2024} }
相关链接
- 论文:https://proceedings.neurips.cc/paper_files/paper/2024/file/21cf8411ed825614e00006a1d9aab7e4-Paper-Conference.pdf
- 项目页面:https://plfb-football.github.io/




