Anglerfish

A chess engine that plays from a learned evaluation, and the trainer that produces it.

The engine is Rust: a UCI front end, a search, and an evaluator the net plugs into. The trainer is Python: it reads the Lichess evaluation dump, turns positions into feature rows, and fits the net the engine loads.

Relies on esca — the chess library both sides read positions through: rules, position facts, PGN, opening books and a UCI client, as the esca crate and the esca Python package.

Layout

pyanglerfish/       trainer, data tooling, CLI                      (Python)
tests/              tests for the Python side
rs_anglerfish/      Cargo workspace                                 (Rust)
  anglerfish-core/  engine: UCI, search, evaluator interface
docs/               architecture and training
data-external/      the Lichess dump; gitignored, symlinked in worktrees

Python

uv sync --all-groups
uv run pytest
uvx ruff check .
uvx ruff format --check .
uvx pyrefly check

Train the net over the Lichess evaluation dump — see docs/training.md:

uv run python -m pyanglerfish.train --help

Rust

cd rs_anglerfish
cargo build --release
cargo test
cargo fmt --check
cargo clippy --all-targets -- -D warnings

The engine reads UCI commands on stdin; add rs_anglerfish/target/release/anglerfish to any chess GUI. Set RUST_LOG=debug for a trace on stderr. Protocol conformance is checked with uci-test-suite:

uvx uci-test-suite ./target/release/anglerfish

Documentation

• docs/architecture.md — the engine, the trainer, and how they are split.
• docs/training.md — the dump pipeline, the net and the checkpoint.

License

MIT — see LICENSE.

Acknowledgements

Lichess, for the evaluation dump and the game database; Stockfish and Leela Chess Zero, the engines this one spars with and is measured against. The library's own credits are in esca's README.