Downloads · 30 days
5
11% of all-time downloads
humanchess-1450-1550-blitz-checkmates
hd787/humanchess-1450-1550-blitz-checkmates
humanchess-1450-1550-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 1450-1550 Blitz Checkmates is a human-move policy model trained from scratch by imitation on Lichess blitz games where both players were rated 1450-1550. The dataset contains complete games ending in checkm…
All-time downloads
46
Public
Repo size
525 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt525 MB · 100%
At a glance
- Task
- Reinforcement Learning
- Library
- pytorch
- License
- mit
- Access
- Public
- Created
- Jul 20, 2026
- Updated
- Jul 20, 2026
- SHA
- 491b4a58
Model card
From the Hugging Face model README
HumanChess 1450-1550 Blitz Checkmates
HumanChess 1450-1550 Blitz Checkmates is a human-move policy model trained from scratch by imitation on Lichess blitz games where both players were rated 1450-1550. The dataset contains complete games ending in checkmate. This is a policy-only network without search or a value head.
This archival release contains the epoch-4 FP32 PyTorch training checkpoint. The INT8 version is intentionally not included because it underperformed the 1250-1350 INT8 model in direct matches. This model is published for continuity of the rating-band series and is not the hosted serving model.
File
| File | Format | Use |
|---|---|---|
checkpoints/cnn-128x6-1450-1550-blitz-checkmates-epoch004.pt | PyTorch FP32 checkpoint | Model, optimizer, and training metadata |
The checkpoint is approximately 525 MB because it includes optimizer state.
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: 4
Validation
| Metric | Value |
|---|---|
| NLL | 1.7226 |
| Perplexity | 5.599 |
| Top-1 accuracy | 46.28% |
| Top-3 accuracy | 74.33% |
| Top-5 accuracy | 84.69% |
In a 30-game candidate round robin, epoch 4 ranked first among epochs 4, 5, 6, and 8, scoring 55%. That small tournament is candidate-selection evidence, not an absolute strength rating.
Inference
human-chess-uci \
--checkpoint checkpoints/cnn-128x6-1450-1550-blitz-checkmates-epoch004.pt \
--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.
- This release contains FP32 only and is not intended as the hosted model.
- Tournament Elo differences are relative, not absolute human or engine ratings.
- Task
- Reinforcement Learning
- Library
- pytorch
- License
- mit
- Created
- Jul 20, 2026
- Updated
- Jul 20, 2026