Marina nano (192x6)
The nano network for Marina,
a Rust UCI chess engine that searches with batched PUCT over a transformer evaluated on
the GPU. Six layers, 192-wide, 6 heads, d_ff 768: 2.74M parameters. Trained on the
engine-all-v2 corpus of Stockfish-labelled positions (side-to-move encoding, policy
and WDL targets, material-aware WDL rescoring), 2.4M steps at batch 512.
| file | what |
|---|---|
model.safetensors, model.json |
the exported network (fp32 weights + architecture card), what Marina loads: setoption name WeightsFile value <this directory> |
vectors.npz |
2,048 golden positions with the reference outputs; marina verify --weights . checks a build reproduces them |
checkpoint-s2400000.pt, train-config.yaml |
the PyTorch training checkpoint and its config, for fine-tuning or continued training with chessformer |
Strength
Stockfish-anchored 120+1 ladder (Stockfish 17.1 UCI_LimitStrength 2600/3000/3190, one
thread, paired-colour 8moves_v3 openings, Python searcher, no tablebases):
3099 ± 24 (510 games). Under Marina the same network scores more per second; its
relative ratings against full-strength engines are kept in the repository README
("Rust engine"). The Stockfish UCI_Elo scale reads about 190 below CCRL Blitz.
Provenance and naming
Training run tiny-engine-192x6-2400k-b512-stm-continue-seed0, checkpoint step
2,400,000. This is the same network that the repository's result tables call
tiny-engine-192x6 / "Tiny 192x6" and that was published under
akhileshsharma99/chessformer-tiny-engine-192x6-3200k; the engine and everything from
here on call it nano. Sizes: nano 192x6 (2.7M), small 512x12 (38M).