onepiece1999/StatePlay-Dataset

Dataset

StatePlay SF3 Dataset

3

44 commits

1 linked in READMEs

updated Jul 31, 2026

See the code

README

StatePlay SF3 Dataset

This repository contains the curated Street Fighter III training data used by StatePlay.

Paper: https://huggingface.co/papers/2607.26754

Code: https://github.com/Jimntu/StatePlay

Dataset contents

The SF3 subset contains 10,000 five-second gameplay clips. Every row in SF3/metadata_state_polish.csv references:

  • video: a gameplay clip at SF3/clips/*/video.mp4;
  • action: an aligned action and game-state table at SF3/clips/*/actions.parquet;
  • prompt: a structured NPC behavior and strategy description;
  • distribution, strategy, and category: auxiliary annotations.

Each Parquet file contains controller inputs together with the StatePlay state variables: timer, both players' health, and both players' meters.

Directory layout

SF3/
├── metadata_state_polish.csv
└── clips/
    └── clip_xxxxxxx/
        ├── video.mp4
        └── actions.parquet

Using the data with StatePlay

After downloading the repository, point StatePlay to its SF3 directory:

export STATEPLAY_DATA_ROOT=/path/to/StatePlay-Dataset/SF3
export STATEPLAY_METADATA="${STATEPLAY_DATA_ROOT}/metadata_state_polish.csv"

Intended use

The intended use is non-commercial academic research on game world models, video generation, action grounding, and game AI.

Third-party intellectual property

Street Fighter III and the game visuals, characters, audio, trademarks, and other game assets appearing in the clips are the property of Capcom Co., Ltd. No ownership of the underlying game intellectual property is claimed or transferred. No game ROMs, BIOS files, or emulator binaries are included.

The CC BY-NC 4.0 license applies only to the original text annotations, action and state traces, dataset organization, and documentation. The underlying game footage is not relicensed. See LICENSE and NOTICE for the full terms.

Maintainer and takedown contact

Zijun Lin
Email: zijun002@e.ntu.edu.sg

Rights holders should contact the maintainer with the affected file paths and a description of the claim. Substantiated requests will be reviewed and handled within a reasonable timeframe.

fighting-game
game-ai
street-fighter
video
world-model

onepiece1999/StatePlay-Dataset

Dataset

StatePlay SF3 Dataset

3

44 commits

1 linked in READMEs

updated Jul 31, 2026

See the code

README

StatePlay SF3 Dataset

This repository contains the curated Street Fighter III training data used by StatePlay.

Paper: https://huggingface.co/papers/2607.26754

Code: https://github.com/Jimntu/StatePlay

Dataset contents

The SF3 subset contains 10,000 five-second gameplay clips. Every row in SF3/metadata_state_polish.csv references:

  • video: a gameplay clip at SF3/clips/*/video.mp4;
  • action: an aligned action and game-state table at SF3/clips/*/actions.parquet;
  • prompt: a structured NPC behavior and strategy description;
  • distribution, strategy, and category: auxiliary annotations.

Each Parquet file contains controller inputs together with the StatePlay state variables: timer, both players' health, and both players' meters.

Directory layout

SF3/
├── metadata_state_polish.csv
└── clips/
    └── clip_xxxxxxx/
        ├── video.mp4
        └── actions.parquet

Using the data with StatePlay

After downloading the repository, point StatePlay to its SF3 directory:

export STATEPLAY_DATA_ROOT=/path/to/StatePlay-Dataset/SF3
export STATEPLAY_METADATA="${STATEPLAY_DATA_ROOT}/metadata_state_polish.csv"

Intended use

The intended use is non-commercial academic research on game world models, video generation, action grounding, and game AI.

Third-party intellectual property

Street Fighter III and the game visuals, characters, audio, trademarks, and other game assets appearing in the clips are the property of Capcom Co., Ltd. No ownership of the underlying game intellectual property is claimed or transferred. No game ROMs, BIOS files, or emulator binaries are included.

The CC BY-NC 4.0 license applies only to the original text annotations, action and state traces, dataset organization, and documentation. The underlying game footage is not relicensed. See LICENSE and NOTICE for the full terms.

Maintainer and takedown contact

Zijun Lin
Email: zijun002@e.ntu.edu.sg

Rights holders should contact the maintainer with the affected file paths and a description of the claim. Substantiated requests will be reviewed and handled within a reasonable timeframe.

fighting-game
game-ai
street-fighter
video
world-model