maia-2200

maia-2200 (Lc0 fine-tune)

• config.yaml: The config file was originally downloaded from: https://github.com/CallOn84/LeelaNets/blob/refs/heads/main/Nets/Maia%202200/Maia-2200.yaml
• The model was originally downloaded from: https://github.com/CallOn84/LeelaNets/raw/refs/heads/main/Nets/Maia%202200/maia-2200.pb.gz

Use with ChessNets

from chessnets.pipelines.lc0 import Lc0Pipeline

pipeline = Lc0Pipeline.from_pretrained("shermansiu/maia-2200")
predictions = pipeline("1. e4 e5")

Checkpoint artifacts

• config.json: canonical, fully materialized ChessNets architecture config
• model.safetensors: decoded float32 PyTorch tensors
• provenance.json: source and conversion hashes
• original/model.pb.gz: original Lc0 protobuf checkpoint
• training/original_lczero_training_config.yaml: original training config

The learned protobuf weight layers used LINEAR16 encoding. Conversion decoded those values into float32 tensors; it cannot restore precision already lost in the source encoding.

Intended use and limitations

This Maia-1 network is intended to imitate moves from its target human rating group. It is a move-prediction model, not a calibrated estimate of an individual player's strength.

Review legal moves and probabilities in context. Outputs are not guarantees of playing strength, fairness, or suitability for high-stakes decisions.

Citation

Please cite the original Maia work:

@inproceedings{mcilroyyoung2020maia,
  title={Aligning Superhuman AI with Human Behavior: Chess as a Model System},
  author={McIlroy-Young, Reid and Sen, Siddhartha and Kleinberg, Jon and Anderson, Ashton},
  booktitle={Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining},
  year={2020},
  doi={10.1145/3394486.3403219}
}

See the official Maia repository and paper.