Downloads · 30 days
4
7% of all-time downloads
humanchess-1250-1350-blitz-checkmates
hd787/humanchess-1250-1350-blitz-checkmates
humanchess-1250-1350-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 1250-1350 Blitz Checkmates is a human-move policy model trained by imitation on Lichess blitz games where both players were rated 1250-1350. The dataset contains complete games ending in checkmate. This is…
All-time downloads
56
Public
Repo size
569 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt525 MB · 92%
At a glance
- Task
- Reinforcement Learning
- Library
- pytorch
- License
- mit
- Access
- Public
- Created
- Jul 14, 2026
- Updated
- Jul 14, 2026
- SHA
- 6c6b276a
Model card
From the Hugging Face model README
HumanChess 1250-1350 Blitz Checkmates
HumanChess 1250-1350 Blitz Checkmates is a human-move policy model trained by imitation on Lichess blitz games where both players were rated 1250-1350. The dataset contains complete games ending in checkmate. This is a policy-only network without search or a value head.
This release contains the epoch-5 FP32 PyTorch training checkpoint and a standalone, statically quantized INT8 TorchScript model for x86 CPU inference.
Files
| File | Format | Use |
|---|---|---|
checkpoints/cnn-128x6-1250-1350-blitz-checkmates-epoch005.pt | PyTorch FP32 checkpoint | Model, optimizer, and training metadata |
artifacts/cnn-128x6-1250-1350-blitz-checkmates-epoch005-int8.ts | INT8 TorchScript | Compact x86 CPU inference model |
The FP32 training checkpoint is approximately 525 MB, including optimizer state. The standalone INT8 artifact is approximately 44 MB.
Architecture
- CNN policy network with 128 channels and 6 residual blocks
- Input:
19 x 8 x 8board tensor - Output: fixed move-vocabulary logits, masked to legal moves at inference
- Saved epoch: 5
Validation
Full epoch-5 validation metrics:
| Metric | Value |
|---|---|
| NLL | 1.7104 |
| Perplexity | 5.531 |
| Top-1 accuracy | 46.78% |
| Top-3 accuracy | 74.73% |
| Top-5 accuracy | 84.91% |
INT8 quantization
On an 8,192-position comparison sample:
| Metric | FP32 | INT8 |
|---|---|---|
| NLL | 1.7352 | 1.7496 |
| Top-1 accuracy | 45.95% | 45.70% |
| Top-3 accuracy | 74.19% | 73.90% |
| Top-5 accuracy | 84.36% | 84.27% |
Raw-logit argmax agreement was 84.74%. In a color-balanced 100-game match at temperature 0.5 with the first six plies sampled, FP32 scored 51%: 44 wins, 42 losses, and 14 draws. The estimated FP32 advantage was 7 Elo with a 64-Elo error margin, making the playing-strength difference statistically indistinguishable from zero.
Inference
Load the FP32 checkpoint with the HumanChess UCI application:
human-chess-uci \
--checkpoint checkpoints/cnn-128x6-1250-1350-blitz-checkmates-epoch005.pt \
--device cpu
Load the standalone INT8 model with --torchscript:
human-chess-uci \
--checkpoint artifacts/cnn-128x6-1250-1350-blitz-checkmates-epoch005-int8.ts \
--torchscript \
--device cpu
Tournament play used --temperature 0.5 --sample-plies 6.
Limitations
- Policy-only: no search and no value evaluation.
- Trained on a narrow rating and time-control distribution.
- The INT8 artifact is intended for x86 CPU execution.
- Tournament Elo differences are relative, not absolute human or engine ratings.
- Task
- Reinforcement Learning
- Library
- pytorch
- License
- mit
- Created
- Jul 14, 2026
- Updated
- Jul 14, 2026