yonghoon96/puzzle-4x4-play-10m-v0

Puzzle-4x4-Play 10M Dataset

This dataset contains 10M transitions for the puzzle-4x4-play task from OGBench. It was used for the experiments in the paper Trust Region Q Adjoint Matching.

Project Page | Code | Paper

Dataset Structure

• 10 files of ~1M transitions each (split for easier loading)
• Each file includes:
  • Training set: 1,000 episodes × 1,001 steps
  • Validation set: 100 episodes × 1,001 steps
• Observation dim: 83, Action dim: 5

Files

puzzle-4x4-play-v0-000.npz / -000-val.npz (seed=0)
puzzle-4x4-play-v0-001.npz / -001-val.npz (seed=1)
...
puzzle-4x4-play-v0-009.npz / -009-val.npz (seed=9)

Usage

You can download and load the dataset using the huggingface_hub library:

import numpy as np
from huggingface_hub import hf_hub_download

# Download a specific file
file_path = hf_hub_download(
    repo_id="yonghoon96/puzzle-4x4-play-10m-v0",
    filename="puzzle-4x4-play-v0-000.npz",
    repo_type="dataset"
)

# Load dataset
data = np.load(file_path)
print(data.files)  # ['observations', 'actions', 'terminals', 'qpos', 'qvel', 'button_states']

Alternatively, to download the entire dataset:

from huggingface_hub import snapshot_download

repo_path = snapshot_download(
    repo_id='yonghoon96/puzzle-4x4-play-10m-v0',
    repo_type='dataset'
)

Generation Details

• Environment: puzzle-4x4-v0
• Dataset Type: play
• Generation Script: OGBench generate_manipspace.py

Citation

@inproceedings{dong2026trqam,
    author    = {Yonghoon Dong and Kyungmin Lee and Changyeon Kim and Jaehyuk Kim and Jinwoo Shin},
    title     = {Trust Region Q Adjoint Matching},
    booktitle = {Advances in Neural Information Processing Systems},
    year      = {2026}
}

@inproceedings{ogbench_park2025,
  title={OGBench: Benchmarking Offline Goal-Conditioned RL},
  author={Park, Seohong and Frans, Kevin and Eysenbach, Benjamin and Levine, Sergey},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2025},
}

License

MIT License (same as OGBench)