humanchess-650-750-blitz

HumanChess Beginner — 650-750 Blitz

HumanChess 650-750 Blitz is a chess move policy model trained by imitation on 650-750 Elo Lichess blitz games. Given a board position, it scores a fixed chess move vocabulary, masks illegal moves, and returns a distribution over legal moves.

For product use, this model is presented as the Beginner tier, roughly a 300-level difficulty. That number is an informal product label based on its observed weakness against the later HumanChess models; it is not a calibrated Chess.com, Lichess, FIDE, or engine Elo rating. The precise provenance remains 650-750-rated Lichess blitz training data.

This is not a search engine and does not include a value head. It is intended to produce low-rated, human-like blitz moves rather than strongest-play chess.

Model

• Architecture: CNN policy
• Channels: 128
• Residual blocks: 6
• Input: 19 x 8 x 8 board tensor
• Output: fixed move-vocabulary logits, masked to legal moves at inference
• Checkpoint: checkpoints/v3-cnn-128x6-20epoch.pt
• FP32 inference model: artifacts/v3-cnn-128x6-20epoch-fp32.ts
• INT8 model: artifacts/v3-cnn-128x6-20epoch-int8.ts
• Saved epoch: 5

Files

The full training checkpoint is approximately 501 MiB, the inference-only FP32 artifact is approximately 167 MiB, and the INT8 artifact is approximately 43 MiB. The FP32 TorchScript export matched the source checkpoint with a maximum absolute output difference of 2.86e-6 on its verification input.

Validation Metrics

Validation metrics from the saved checkpoint:

INT8 Quantization

The INT8 version uses static x86 quantization. Relative to FP32, validation NLL increased by approximately 0.047 and top-1 accuracy decreased by approximately 1.59 percentage points.

In a color-balanced 100-game match, FP32 scored 59% with 34 wins, 16 losses, and 50 draws. Its estimated advantage was 63 Elo with a 48-Elo error margin. The INT8 model is therefore measurably weaker, but retains the intended beginner-tier behavior while being much smaller and faster on x86 CPUs.

Intended Use

This model is useful for experiments with human-like chess move prediction, training tools, casual chess bots, and model-vs-model comparisons. It should not be treated as a calibrated Elo engine or a tactical oracle.

Usage

Use the inference provider repository:

git clone https://github.com/hd787/ChessModelServer.git
cd ChessModelServer

Then download this checkpoint from Hugging Face and point the server at:

checkpoints/v3-cnn-128x6-20epoch.pt

For the INT8 artifact, use the TorchScript loader:

human-chess-uci \
  --checkpoint artifacts/v3-cnn-128x6-20epoch-int8.ts \
  --torchscript \
  --device cpu

The FP32 inference artifact uses the same TorchScript loader:

human-chess-uci \
  --checkpoint artifacts/v3-cnn-128x6-20epoch-fp32.ts \
  --torchscript \
  --device cpu

Training Data

The checkpoint was trained from a v3 split of Lichess blitz games filtered to roughly 650-750 Elo players. The model is supervised by the human move played in each retained position.

Limitations

• Policy-only: no search and no value evaluation.
• Move quality is intentionally limited by the training distribution.
• The model can choose illegal-looking strategic plans even though individual moves are masked to legal moves.
• Tactical puzzle fine-tuning experiments degraded tournament performance in our launch-style tests, so this release uses the blitz imitation checkpoint.
• The approximate 300-level product difficulty is deliberately not a platform rating claim and may feel different across opponents and playing conditions.