p-doom/atari-assault-dataset

Atari-Assault Dataset

This is a large dataset of 10M video frames and actions collected from the Assault atari environment (Bellemare et al., 2012) in order to train world models.
The dataset enables reproducible, large-scale experiments in action-conditioned video prediction. It is meant to be used with Jasmine, our JAX-based world modeling codebase.

---

Dataset Summary

• Environment: Atari Learning Environment
• Frames: 10 million
• Resolution: 84 × 84
• Format: ArrayRecord (for fast I/O)
• Splits: train / val / test
• License: CC0 1.0

---

Usage

This dataset is part of the Jasmine repository release. Frames were collected during Rainbow (Hessel et al. 2017) agent training. You can find the dataset generation code at https://github.com/p-doom/jasmine/tree/main/data/jasmine_data.

The ArrayRecord format enables efficient dataloading using Grain and is optimized for the Jasmine dataloader.

You can download the dataset using the huggingface-cli tool.

huggingface-cli download --repo-type dataset p-doom/atari-assault-dataset --local-dir <data_path>

To start a training run using Jasmine, simply pass the train and val split to the training script.

python jasmine/baselines/maskgit/train_tokenizer_vqvae.py \
    --data_dir <data_path>/train \
    --val_data_dir <data_path>/val \
    ...

Citation

If you use our Atari datasets, please cite our work:

@article{
    mahajan2025jasmine,
    title={Jasmine: A simple, performant and scalable JAX-based world modeling codebase},
    author={Mihir Mahajan and Alfred Nguyen and Franz Srambical and Stefan Bauer},
    journal = {p(doom) blog},
    year={2025},
    url={https://pdoom.org/jasmine.html},
    note = {https://pdoom.org/blog.html}
}