johanventa/gm_and_random_chess

Datasets used in project.

Random-move dataset:

The dataset are 5000 chessgames datasets, with random-moves, and their state-values, based on Stockfish-18 evaluations added to the board. The data is encoded into 17-state per board position FEN embeddings, and into 112-state per position data.

The file is called rm_data.h5

GrandMaster data (GM)

The Second, called "50k_gm_dataset.h5" are 50k games, around 123 million legal moves, and 4,150,720 positions, embedded in PGN. The 20k data is taken from human GM data from Lichess, and only forms a small subsection of this Kaggle set, with a total dataset size of (n= 4,193,152) The original data is in PGN format. This dataset is in FEN + memory states. The Data also has move values for search depth-9 for all legal moves from the positions, judged by a Stockfish-18 engine.

The 50 000 games are converted into 17-state FEN positions and into the 112-state per position data.

Second GM data

The second GM dataset found in "state_gm_dataset.h5", which contains 700,000 games and 58,289,358 positions, with only the move based state-values judged by the Stockfish engine. This means that the stockfish engine does not evaluate all the possible/legal moves with search dept-9 like the full gm-dataset (50k_gm_dataset.h5). This is done to significantly speed up the conversion from original PGN to states useable by ResNet, Transformer and ChessFormers-state-value head.

Creators:

Johan Baumann, johanbau@chalmer.se

Rasmus Hulthe, rasmushu@chalmers.se