ChessModels stores chess neural-network weight files and minimal documentation.

The repository is intentionally narrow in scope: it exists to keep large model binaries, hashes, and lightweight metadata in one place.

• Versioned chess neural-network weight files
• Stable download targets for reproducible setups
• Lightweight metadata for the files in models/

• A general-purpose training codebase
• A dataset repository
• A general chess engine toolkit

The diagrams below are rendered from Graphviz .dot sources in assets/.

Source: assets/lc0-610153-architecture.dot

Source: assets/lc0-744706-architecture.dot

assets/   banner and repository artwork
models/   model binaries and per-file notes
training/ puzzle-classifier training scripts

Clone or download this repository, then use the file you want from models/.

If you need the per-model notes, checksums, or provenance that has been documented so far, see models/README.md.

Run the default puzzle-classifier training pipeline:

./train.sh

The default pipeline scans the USB stack data, writes a puzzle-focused local dataset under training/classifier/data/, trains the classifier CNN with CUDA when available, and exports the best checkpoint to models/.