SC2EGSet: StarCraft II Esport Replay and Game-state Dataset

Area of Science:

• Computer Science
• Sports Science
• Data Science

Background:

• Esports, a rapidly growing field, presents unique opportunities for data-driven research due to its inherent data richness.
• Existing research in esports is limited by the availability of comprehensive, structured datasets suitable for advanced analysis.

Purpose of the Study:

• To provide a large-scale, publicly accessible dataset of StarCraft II esports matches for scientific inquiry.
• To facilitate the use of esports data in statistical and machine learning (ML) modeling, comparable to laboratory measurements.
• To share open-source tools developed for data processing and analysis, supporting broader adoption in scientific research.

Main Methods:

• Collected and processed game-state information from 17,930 StarCraft II replay files across major tournaments since 2016.
• Developed and open-sourced custom tools, including PyTorch and PyTorch Lightning API abstractions, for data loading and modeling.
• Organized data into 55 "replaypacks" for ease of access and utilization.

Main Results:

• A substantial, publicly available dataset of StarCraft II esports data has been curated.
• The dataset includes raw and pre-processed game-state information, offering a valuable resource for researchers.
• Associated open-source software tools are provided to aid in data analysis and model development.

Conclusions:

• The released dataset and tools significantly lower the barrier for scientific research in esports.
• This resource is expected to spur advancements in Artificial Intelligence (AI), ML, psychology, Human-Computer Interaction (HCI), and sports analytics.
• The availability of high-quality esports data enables novel research avenues in both supervised and self-supervised learning tasks.