metadata

license: mit
datasets:
  - Xmaster6y/stockfish-debug
name: Xmaster6y/gpt2-stockfish-debug
results:
  - task: train
    metrics:
      - name: train-loss
        type: loss
        value: 0.151
        verified: false
      - name: eval-loss
        type: loss
        value: 0.138
        verified: false

Model Card for gpt2-stockfish-debug

Training Details

The model was trained during 1 epoch on the Xmaster6y/stockfish-debug dataset (no hyperparameter tuning done). The samples are:

{"prompt":"FEN: {fen}\nMOVE:", "completion": " {move}"}

Two possible simple extensions:

Expand the FEN string: r2qk3/... -> r11qk111/... or equivalent
Condition with the result (ELO not available in the dataset):

{"prompt":"RES: {res}\nFEN: {fen}\nMOVE:", "completion": " {move}"}

Use the Model

import chess
from transformers import AutoModelForCausalLM, AutoTokenizer

def next_move(model, tokenizer, fen):
    input_ids = tokenizer(f"FEN: {fen}\nMOVE:", return_tensors="pt")
    input_ids = {k:v.to(model.device) for k,v in input_ids.items()}
    out = model.generate(
        **input_ids, 
        max_new_tokens=10,
        pad_token_id=tokenizer.eos_token_id,
        do_sample=True,
        temperature=0.1,
        )
    out_str = tokenizer.batch_decode(out)[0]
    return out_str.split('MOVE:')[-1].replace("<|endoftext|>", "").strip()

board = chess.Board()
model = AutoModelForCausalLM.from_pretrained('Xmaster6y/gpt2-stockfish-debug')
tokenizer = AutoTokenizer.from_pretrained('gpt2')
tokenizer.pad_token = tokenizer.eos_token
for i in range(100):
    fen = board.fen()
    move_uci = next_move(model, tokenizer, fen)
    try:
        print(move_uci)
        move = chess.Move.from_uci(move_uci)
        if move not in board.legal_moves:
            raise chess.IllegalMoveError
        board.push(move)
    except chess.IllegalMoveError:
        print(board)
        print("Illegal move", i)
        break