Edit model card

Model Card for gpt2-stockfish-debug

See my blog post for additional details.

Training Details

The model was trained during 1 epoch on the yp-edu/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

The following code requires python-chess (in addition to transformers) which you can install using pip install python-chess.

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("yp-edu/gpt2-stockfish-debug")
tokenizer = AutoTokenizer.from_pretrained("yp-edu/gpt2-stockfish-debug")  # or "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)
        outcome = board.outcome()
        if outcome is not None:
            print(board)
            print(outcome.result())
            break
    except chess.IllegalMoveError:
        print(board)
        print("Illegal move", i)
        break
else:
    print(board)
Downloads last month
37
Safetensors
Model size
124M params
Tensor type
F32
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Dataset used to train yp-edu/gpt2-stockfish-debug

Space using yp-edu/gpt2-stockfish-debug 1