maxmill/aq-mario-smb1

AQ-Mario SMB1 Dataset

This is the 8,000-episode Super Mario Bros. 1-1 dataset used to train the AQ-Mario EMA-JEPA world model. The source code, shard loader, training recipe, and evaluation protocol are in the AQ-Mario GitHub repository.

Contents

Track	Episodes	Observations	Files
random	4,000	2,674,774	256 compressed .npz shards
ppo	4,000	2,202,594	256 compressed .npz shards
Total	8,000	4,877,368	512 shards, about 27.4 GB

Each shard contains flat arrays and episode offsets. The exact schema and action/frame alignment are documented in data/README.md. The repository is intentionally kept as raw compressed NumPy shards so the dataset can be streamed or consumed by the project's ShardWindows loader.

Dataset snapshots

These previews are generated from representative frames in the local sample shard. They show the raw observations and are included only as visual inspection aids; the authoritative dataset is the complete set of .npz shards.

Download

huggingface-cli download maxmill/aq-mario-smb1 random --repo-type dataset --local-dir ./aq-mario-data/random
huggingface-cli download maxmill/aq-mario-smb1 ppo --repo-type dataset --local-dir ./aq-mario-data/ppo

The model and dataset are separate repositories so users can download only the artifact they need. The model page links back here, and the GitHub project links to both.

Data and rights

These are derived gameplay observations from Super Mario Bros. 1-1. AQ-Mario is an independent research project and is not affiliated with or endorsed by Nintendo. Use of the derived data is subject to the rights and terms applicable to the underlying game content.