Repository with code for training a LLM with Leela Chess Zero embeddings to explain Chesspositions. By FHNW University of Applied Sciences and Arts, Institute of Data Science, Windisch
You can find used data here: https://doi.org/10.5281/zenodo.20666580
Repository for training a language-model with chess embeddings from Leela Chess Zero. The Goal is to get the LLM to understand Leela Chess Zero Embeddings and then use them to explain Positions for Humans.
Repository with inference code for Leela Chess Zero used to generate the datasets to train the LLMs. This Repository contains scripts to generate new datasets in form of Chess Position Embeddings from Leela and new generated Text descriptions for the positions. This includes text descriptions with position features gathered from Stockfish 11 or also LLM generated Questions for these descriptions! For data generation scripts you need to have a Stockfish model and an onnx Version of Leela like BT4.
Preprocessing scripts for scraped chess and text data. The preprocessing includes converting the data to the format that Leela Chess Zero uses for its input and then generating the embeddings for the positions.
(proof of concept)
Repository to predict if a Text matches to its chess embedding from leela. Used as a Proof of concept if a model is able to understand if a text matches to the Leela embedding. Some of the text and embedding pairs are real some are shuffled. The model has to predict 1 or 0.
This work was supported by the Hasler Foundation (Hasler Stiftung) under grant no. 2025-05-05-522
