Marina small (512x12)
The small network for Marina,
a Rust UCI chess engine that searches with batched PUCT over a transformer evaluated on
the GPU. Twelve layers, 512-wide, 16 heads, d_ff 2048: 38.2M parameters. Trained on
the engine-all-v2 corpus of Stockfish-labelled positions (side-to-move encoding, policy
and WDL targets, Syzygy 6-piece rescoring, material-aware WDL rescoring), 4.8M steps.
| 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-s4800000.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):
3191 ± 21 (510 games), the strongest network in the project so far — +92 over
nano at the same clock while searching about 60% of the nodes. Under Marina it has not
been tuned yet (its forward is ~14× the cost of nano's). The Stockfish UCI_Elo scale
reads about 190 below CCRL Blitz.
Provenance and naming
Training run scaled-engine-512x12-4800k-stm-tb6-seed0-r2, checkpoint step 4,800,000.
This is the network the repository's result tables call scaled-engine-512x12 /
"Scaled 512x12" (tb6-s4800000); the engine and everything from here on call it
small. Sizes: nano 192x6 (2.7M), small 512x12 (38M).