Starcraft AI Research Dataset

1 week after



We release the largest StarCraft: Brood War replay dataset yet, with 65646 games. The full dataset after compression is 365 GB, 1535 million frames, and 496 million player actions. The entire frame data was dumped out at 8 frames per second. We made a big effort to ensure this dataset is clean and has mostly high quality replays. You can access it with TorchCraft in C++, Python, and Lua. The replays are in an AWS S3 bucket at s3://stardata. Read below for more details, or our whitepaper on arXiv for more details.

Installing TorchCraft

Note: The current set of replays are only compatible with the 1.3.0 version of torchcraft included here.

Simply do

git submodule update --init
cd TorchCraft
pip install .

More documentation can be found at Realistically, you will only need the replayer modules, which means you can ignore most of the connecting to starcraft parts. Check out the code to document its use

Note: Please make sure you have libzstd-1.1.4+, torchcraft will compile without it but won't be able to read the replays.

Downloading the Data

You can find the replays in an AWS S3 bucket at s3://stardata

  • s3://stardata/dumped_replays contains the replays in a format readable by TorchCraft
  • s3://stardata/battles are text files, containing one battle each. Each battle is 3 lines:
    • xmin, xmax, ymin, ymax, tmin, tmax: the bounding rectangle for the battle. Multiply time by 3 to get real frame count, or don't to index directly into the dumped datasets.
    • Type and number of units on team 1
    • Type and number of units on team 2
  • s3://stardata/original_replays.tar contains the original replays.

Reproducing Results

Some of the reproduction scripts are included, others scripts will be added as soon as we clean up the code and make it easy to install/run. Simply make and you're good to go. All cpp files can be run like script /path/to/replays/**/*.rep

  • extract_stats tells you some stats about the replays
  • extract_units preprocesses for battle clustering
  • get_corrupt_replays tells you what replays are considered corrupt
  • can be run on the output of extract_units to do battle clustering.


The white paper for the dataset is at:

Lin, Z., G., Jonas, K., Vasil, Synnaeve, G., AIIDE 2017. STARDATA: A StarCraft AI Research Dataset (arxiv)

We attribute most of the replays to bwrep and G. Synnaeve, P. Bessiere, A Dataset for StarCraft AI & an Example of Armies Clustering, 2012. Please see the paper for a complete list of references.


StarData is BSD-licensed. We also provide an additional patent grant.