This repository provides StarData, a large-scale dataset of 65,646 StarCraft: Brood War replays, aimed at AI researchers and practitioners. It enables high-quality analysis and development of AI agents for real-time strategy games, offering 365 GB of compressed data with 1.5 billion frames and 496 million player actions, all captured at 8 frames per second.

How It Works

The dataset is designed for use with TorchCraft, a framework that allows interaction with StarCraft: Brood War replays. TorchCraft provides replayer modules in C++, Python, and Lua, enabling efficient parsing and access to frame-by-frame game data. This approach facilitates detailed analysis of game states, player actions, and unit behaviors, crucial for training and evaluating AI models.

Quick Start & Requirements

• Install TorchCraft: git submodule update --init && cd TorchCraft && pip install .
• Prerequisites: libzstd-1.1.4+ is required for replay parsing.
• Data Access: Replays are available via AWS S3 at s3://stardata or through provided chunked download links. Standardized train, validation, and test sets are also available.
• Documentation: https://github.com/TorchCraft/TorchCraft

Highlighted Details

• Largest StarCraft: Brood War replay dataset available (65,646 games).
• Data captured at 8 frames per second for granular analysis.
• Includes tools for preprocessing, clustering, and reproducing results.
• Compatible with TorchCraft versions 1.3.0 and later.

Maintenance & Community

The project is associated with research from authors like Z. Lin, G. Synnaeve, and others, with a whitepaper available on arXiv. Specific community channels are not explicitly mentioned in the README.

Licensing & Compatibility

StarData is BSD-licensed with an additional patent grant, permitting commercial use and integration with closed-source projects.

Limitations & Caveats

The provided replay data is specifically compatible with TorchCraft version 1.3.0. Some reproduction scripts are still pending cleanup and easier installation.