|
|
--- |
|
|
license: mit |
|
|
tags: |
|
|
- chess |
|
|
- game-ai |
|
|
- pytorch |
|
|
- safetensors |
|
|
library_name: transformers |
|
|
--- |
|
|
|
|
|
# ChessLC0 Chess Model |
|
|
|
|
|
This is BT4, the model behind LeelaChessZero Engine, one of the best Neural Network based engine available. This model is way worse than stockfish but constitute one of the best 0 search heuristics out there. |
|
|
For stronger play, reducing temperature T (lower is stronger) is suggested. |
|
|
|
|
|
## Model Description |
|
|
|
|
|
The ChessLC0 model is a transformer-based architecture designed for chess gameplay. It can: |
|
|
- Predict the next best move given a move history (requires 7 prior boards) |
|
|
- Evaluate chess positions |
|
|
- Generate move probabilities |
|
|
|
|
|
**Important**: This model requires move history (7 prior boards) to work properly. You must provide a list of UCI moves representing the game history. |
|
|
|
|
|
## Please Like if this model is useful to you :) |
|
|
|
|
|
A like goes a long way ! |
|
|
|
|
|
## Usage |
|
|
|
|
|
```python |
|
|
import torch |
|
|
from transformers import AutoModel |
|
|
|
|
|
model = AutoModel.from_pretrained("Maxlegrec/ChessLC0", trust_remote_code=True) |
|
|
device = "cuda" if torch.cuda.is_available() else "cpu" |
|
|
model = model.to(device) |
|
|
|
|
|
# Example usage with move history (model requires 7 prior boards) |
|
|
# This sequence provides enough history for the model |
|
|
move_history = [ |
|
|
"e2e4", "e7e5", "g1f3", "b8c6", "f1b5", "a7a6", "b5a4", |
|
|
"g8f6", "e1g1", "f8e7", "f1e1", "b7b5", "a4b3", "d7d6" |
|
|
] |
|
|
|
|
|
# Sample move from policy |
|
|
move = model.get_move_from_history(move_history, T=0.1, device=device) |
|
|
print(f"Policy-based move: {move}") |
|
|
|
|
|
# Get the best move using value analysis |
|
|
value_move, value = model.get_best_move_value(move_history, T=0, device=device) |
|
|
print(f"Value-based move: {value_move}") |
|
|
print(f"Position value [black_win, draw, white_win]: {value}") |
|
|
|
|
|
# Get position evaluation |
|
|
position_value = model.get_position_value(move_history, device=device) |
|
|
print(f"Position value [current_side_win, draw, opposite_side_win]: {position_value}") |
|
|
|
|
|
# Get move probabilities |
|
|
probs = model.get_move_from_history(move_history, T=1, device=device, return_probs=True) |
|
|
top_moves = sorted(probs.items(), key=lambda x: x[1], reverse=True)[:5] |
|
|
print("Top 5 moves:") |
|
|
for move, prob in top_moves: |
|
|
print(f" {move}: {prob:.4f}") |
|
|
``` |
|
|
|
|
|
## Requirements |
|
|
|
|
|
python-version >=3.13 |
|
|
cuda-version < 13.0 |
|
|
|
|
|
- torch>=2.0.0 |
|
|
- transformers>=4.48.1 |
|
|
- bulletchess>=0.4.0 |
|
|
- numpy>=1.21.0 |
|
|
|
|
|
## Model Architecture |
|
|
|
|
|
- **Transformer layers**: 15 |
|
|
- **Hidden size**: 1024 |
|
|
- **Feed-forward size**: 1536 |
|
|
- **Attention heads**: 32 |
|
|
- **Vocabulary size**: 1858 (chess moves) |
|
|
|
