README.md

metadata

title: README
emoji: 🐢
colorFrom: indigo
colorTo: yellow
sdk: gradio
pinned: false

Leela Chess Zero

All models here are "Chessformers" — encoder-only transformers with the Geometric Attention Bias described in the paper Chessformer: A Unified Architecture for Chess Modeling.

They are expressed in the internal Lc0 format built on Protobuf (hence .pb.gz), which is compatible with the training code and the engine. They can be converted to ONNX through the leela2onnx command through the Lc0 binary, which we recommend for mechanistic interpretability research.

BT4 is our largest model at 190M and corresponds to Leela-CF. BT4 and BT3 were trained with SFT on a prior RL run, and the T1 models were distilled from a larger transformer. The models range in skill from expert-level (T1-10L-256D) to grandmaster-plus-level (BT4) at Blitz time controls with greedy sampling of policy prediction.

See also the website, Discord server, and a longest model list of the Leela Chess Zero project.