The EsportsBench datasets are meant to facilitate research and experimentation on real world competition data spanning many years of competitions including a diverse range of genres and competition formats.
This code is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
To collect the data yourself, you will need to obtain Aliculac and Liquipedia LPDB API key. Aligulac keys can be generated at http://aligulac.com/about/api/ and Liquipedia LPDB keys can be requested in the Liquipedia discord server.
Add your key(s) to .dotenv-template
and rename it to .env
so that the data pipelines can access them.
To exactly reproduce the results of Table 2 of the paper, follow these steps exactly:
Clone the repo and enter:
git clone https://github.com/cthorrez/esports-bench
cd esports-bench
Checkout the v0.0.1 tag
git checkout tags/v0.0.1
Create conda env
conda create -n esportsbench python=3.11
Activate env
conda activate esportsbench
Install requirements
pip install riix==0.0.3
pip install -e .
Download the data from HuggingFace
cd esportsbench
python data_pipeline/get_data_from_hf.py
Run the broad hyperparameter sweep
python experiments/broad_sweep.py --config_file experiments/configs/broad_sweep_config.yaml
Run the fine hyperparameter sweep
python experiments/fine_sweep.py --config_dir experiments/sweep_results/broad_sweep_7D_1000
Run the evaluation script using the best parameters
python eval/bench.py --config_dir experiments/sweep_results/fine_sweep_7D_1000/
Note each sweep will take many hours, it takes 5 hours on my desktop and 17 on my laptop.
You can use the -np <num_cores>
flag to increase the number of processes utilized to be the number of cores you have
The data collected by these pipelines is collected from different sources with their own licenses. If you reproduce the the data collection and experiments understand the retrieved data and results falls under those licenses.
The StarCraft II data is from Aligulac
The League of Legends data is from Leaguepedia under a CC BY-SA 3.0
The data for all other games is from Liquipedia under a CC BY-SA 3.0