
PLAICraft Dataset
This dataset contains approximately 200 hours of Minecraft agent gameplay from the PLAICraft project. It contains both the training and validation sets, the exact separation can be found in the global metadata tables provided below. It downloads as a .zip file from Amazon S3 and includes a data/ folder with both raw and encoded data.
Learn More About the Data
This data was collected across three anonymous players (Dante, Morgan, and Xander). They played together as a team in the same player base throughout their time on the server, interacting constantly in most of their gameplay sessions. Thus, this subset is a perfect example of the rich social dynamics in our dataset. Dante self-identifies as a 49-year-old male player with “Regular” level experience, Morgan self-identifies as a 9-year-old male player with “Pro” level experience and Xander self-identifies as a 37-year-old female player with “Amateur” level experience. Within their 200 hours of gameplay, they performed a wide range of activities, including various types of team collaborations like base construction, biomes exploration, combating each other, building iron farms, etc.




Full Dataset — Direct Download (CLI)
Warning: this dataset is 621 GB.
Metadata
Below is the metadata for both the training & validation sets stored as sqlite3 databases, available as either a .zip file or CLI direct download from Amazon S3.
Metadata — Direct Download (CLI)
Croissant Files
Below are the croissant files for the 200-hour dataset and the full 10,000 hour dataset (not available yet). They download as JSON-LD files from Amazon S3.







Leave a comment