You can deploy this model and start earning money today!

hd787/
humanchess-1550-1650-blitz-checkmates

Name: humanchess-1550-1650-blitz-checkmates
Availability: InStock

reinforcement-learning

mit

Views

No views yet

HumanChess 1550-1650 Blitz Checkmates

HumanChess 1550-1650 Blitz Checkmates is a human-move policy model trained from scratch on an 80 GiB dataset of Lichess blitz games where both players were rated 1550-1650 and the game ended in checkmate. It is a policy-only network without search or a value head.

Epoch 7 was selected independently in FP32 and INT8 gameplay tests. This model is intended for serving as the platform level immediately above the 1250-1350 model. The rating range describes the training data, not an absolute engine Elo.

Files

File	Format	Use
`checkpoints/cnn-128x6-1550-1650-blitz-checkmates-epoch007.pt`	PyTorch FP32 checkpoint	Model, optimizer, and training metadata
`artifacts/cnn-128x6-1550-1650-blitz-checkmates-epoch007-int8.ts`	INT8 TorchScript	Compact x86 CPU serving model

Architecture and validation

CNN policy network: 128 channels, 6 residual blocks
Input: 19 x 8 x 8; output: fixed move-vocabulary logits masked to legal moves
Validation NLL: 1.668047
Top-1 / top-3 / top-5: 47.82% / 75.84% / 85.90%
INT8 calibration check: NLL 1.7705, top-1 45.31%, FP32/INT8 argmax agreement 68.57%

Selection evidence

All matches used CPU inference, two seconds per move, temperature 0.5, and sampling for the first six plies.

FP32 epoch round robin: epoch 7 scored 57.7% over 150 games.
INT8 epoch 7 vs epoch 8: 28-14-58 (57.0%, +49 Elo relative).
FP32 epoch 7 vs 1250 epoch 5: 48-32-20 (58.0%).
FP32 epoch 7 vs 1450 epoch 7: 47-32-21 (57.5%).
INT8 epoch 7 vs 1250 epoch 5: 39-41-20 (49.0%; effectively tied).
INT8 epoch 7 vs 1450 epoch 8: 56-32-12 (62.0%).

Inference

bash

human-chess-uci \
  --checkpoint artifacts/cnn-128x6-1550-1650-blitz-checkmates-epoch007-int8.ts \
  --torchscript --device cpu \
  --temperature 0.5 --sample-plies 6

Limitations

Policy-only: no search and no value evaluation.
Trained on a narrow rating, time-control, and checkmate-only distribution.
Tournament Elo differences are relative and have substantial sampling error.
INT8 behavior is not strictly transitive across rating bands.

hd787/humanchess-1550-1650-blitz-checkmates