For an explanation of this project and the models trained for it, please see the Report.
The root folder contains scripts for dataset preprocessing.
chess-mamba-vs-xformer contains the training scripts.
Config files, used to set model configuration and training hyperameters, are in chess-mamba-vs-xformer/config.
Model checkpoints are in chess-mamba-vs-xformer/out. The last checkpoint for completed models (e.g. Mamba and Transformer 50M) are .../anneal/anneal_complete.pt.
chess-gpt-eval has the scripts for model evaluation - playings games against Stockfish or lc0 chess engines. The logs folder contains raw evaluation metrics.
chess-gpt-eval-contrastive likewise has the scripts for model evaluation, but modified for training and evaluation of contrastive activation and linear probes. The logs folder again contains raw evaluation metrics.