HumanChess 1750-1850 Blitz Checkmates
HumanChess 1750-1850 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 1750-1850 and the game ended in checkmate. It is a policy-only network without search or a value head. The rating range describes the training data, not an absolute engine Elo.
Epoch 8 is packaged in both FP32 and INT8. It was selected for cross-band continuity against the 1550-1650 epoch-7 model. Epoch 5 won the broad INT8 internal round robin, but epoch 8 performed substantially better in the cross-band test.
Files
| File | Format | Use |
|---|---|---|
checkpoints/cnn-128x6-1750-1850-blitz-checkmates-epoch008.pt | PyTorch FP32 checkpoint | Model, optimizer, and training metadata |
artifacts/cnn-128x6-1750-1850-blitz-checkmates-epoch008-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 - Epoch-8 validation NLL: 1.668631
- Top-1 / top-3 / top-5: 47.49% / 75.82% / 86.04%
- INT8 calibration check: NLL 1.668020, top-1 46.91%, FP32/INT8 argmax agreement 85.91%
Selection evidence
All matches used CPU inference, two seconds per move, temperature 0.5, sampling for the first six plies, and paired colors.
- FP32 internal round robin: epoch 8 scored 60.7% over 150 games.
- INT8 internal round robin: epoch 8 scored 47.3%; epoch 5 led at 60.7%.
- FP32 epoch 8 vs 1550-1650 epoch 7: 47-30-23 (58.5%, +60 Elo).
- INT8 epoch 8 vs 1550-1650 epoch 7: 75-11-14 (82.0%, +263 Elo).
- INT8 epoch 5 vs 1550-1650 epoch 7: 47-13-40 (67.0%, +123 Elo).
Inference
human-chess-uci \
--checkpoint artifacts/cnn-128x6-1750-1850-blitz-checkmates-epoch008-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 matchup-dependent and not strictly transitive.