Marvin

[GitHub]

Marvin is a human-like chess neural network that models human play at a specified Elo rating and time control. It comes in three sizes: large, small, and tiny.

Model Sizes

Size	Parameters
Large	~49M
Small	~23M
Tiny	~5M

Weight Formats

.onnx - Full ONNX weights

These files are used for inference in the repo.

large/large.onnx
small/marvin_small.onnx
tiny/marvin_tiny.onnx

.pb.gz - Leela-compatible weights (ONNX backend)

These are frozen weights compatible with Leela Chess Zero (lc0), intended for use with the ONNX backend (onnx-cuda or onnx-cpu).

• The time control input is frozen to 300+0 (5 minutes, no increment).
• The Elo input is frozen to a separate weight file for each Elo rating.
• Elo ratings available: 1200-2400

To use with lc0, pass the .pb.gz file as the weights argument with the ONNX backend:

lc0 --backend=onnx-cuda --weights=large/large_1500.pb.gz

Training Checkpoints

Training checkpoints at 750k steps are included for each model size in .pt format.

License

CC BY-SA 4.0