This web content is about the evaluation of LLMs (Language Model Machines) trained on PGN format chess games. Similar to the Othello World paper, this repository allows for the training, evaluation, and visualization of linear probes on LLMs that have been trained to play chess using PGN strings. The author provides examples of visualizations that show where the model “thinks” the white pawns are located. They also trained linear probes on the model’s ability to estimate player ELO while predicting the next character. The web content includes setup instructions and references to additional resources for further information.
https://github.com/adamkarvonen/chess_llm_interpretability