Downloads · 30 days
6
11% of all-time downloads
hd787/humanchess-900-1000-blitz-checkmates
humanchess-900-1000-blitz-checkmates is a reinforcement learning model from hd787. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. It is set up for pytorch. The card lists the license as mit.
HumanChess 900-1000 Blitz Checkmates is a human-move policy model trained by imitation on Lichess blitz games where both players were rated 900-1000. The dataset includes complete games that ended in checkmate; the mo…
Downloads · 30 days
6
11% of all-time downloads
All-time downloads
56
Public
Repo size
569 MB
Likes
0
Public
Click a slice to open those files.
.pt525 MB · 92%
From the Hugging Face model README
HumanChess 900-1000 Blitz Checkmates is a human-move policy model trained by imitation on Lichess blitz games where both players were rated 900-1000. The dataset includes complete games that ended in checkmate; the model itself is a policy network without search or a value head.
This release contains both the original PyTorch training checkpoint and a standalone INT8 TorchScript model optimized for x86 CPU inference.
| File | Format | Use |
|---|---|---|
checkpoints/cnn-128x6-900-1000-blitz-checkmates.pt | PyTorch checkpoint, FP32 | Original model plus optimizer and training metadata |
artifacts/cnn-128x6-900-1000-blitz-checkmates-int8.ts | INT8 TorchScript | Compact CPU inference model |
The FP32 checkpoint is about 525 MB because it includes AdamW optimizer state. Its raw model weights occupy about 175 MB. The INT8 artifact is about 44 MB.
19 x 8 x 8 board tensorMeasured on 300,000 held-out positions:
| Metric | Value |
|---|---|
| NLL | 1.7372 |
| Entropy | 1.6193 |
| Top-1 accuracy | 46.45% |
| Top-3 accuracy | 74.04% |
| Top-5 accuracy | 84.17% |
The INT8 model uses statically quantized x86 convolution and linear layers. On an 8,192-position comparison sample:
| Metric | FP32 | INT8 |
|---|---|---|
| NLL | 1.7209 | 1.7782 |
| Top-1 accuracy | 46.70% | 44.85% |
| Top-3 accuracy | 74.34% | 72.79% |
| Top-5 accuracy | 84.57% | 83.63% |
In a color-balanced 100-game match at temperature 0.5 with the first six plies sampled, FP32 scored 55.5% (50 wins, 39 losses, 11 draws). The estimated FP32 advantage was 38 Elo with a wide 65-Elo error margin, so the playing-strength difference was not statistically conclusive.
With the HumanChess UCI application, load the FP32 checkpoint normally:
human-chess-uci \
--checkpoint checkpoints/cnn-128x6-900-1000-blitz-checkmates.pt \
--device cpu
Load the standalone INT8 model with --torchscript:
human-chess-uci \
--checkpoint artifacts/cnn-128x6-900-1000-blitz-checkmates-int8.ts \
--torchscript \
--device cpu
The tournament configuration used --temperature 0.5 --sample-plies 6.