tokenizer.py

from __future__ import annotations
import json
import os
from typing import Dict, List, Optional
from transformers import PreTrainedTokenizer

class ChessTokenizer(PreTrainedTokenizer):
    model_input_names = ["input_ids", "attention_mask"]
    
    # Special tokens
    PAD_TOKEN = "[PAD]"
    BOS_TOKEN = "[BOS]" 
    EOS_TOKEN = "[EOS]" 
    UNK_TOKEN = "[UNK]"

    # REQUIRED for .from_pretrained to work automatically
    vocab_files_names = {
        "vocab_file": "vocab.json",
    }
    
    def __init__(
        self,
        vocab_file: Optional[str] = None,
        vocab: Optional[Dict[str, int]] = None,
        **kwargs,
    ):
        # Clean kwargs to avoid conflicts
        kwargs.pop("pad_token", None)
        kwargs.pop("bos_token", None)
        kwargs.pop("eos_token", None)
        kwargs.pop("unk_token", None)

        self.vocab_file = vocab_file
        
        # Load vocab
        if vocab is not None:
            self._vocab = vocab
        elif vocab_file is not None and os.path.exists(vocab_file):
            with open(vocab_file, "r", encoding="utf-8") as f:
                self._vocab = json.load(f)
        else:
            self._vocab = self._create_default_vocab()
            
        self._ids_to_tokens = {v: k for k, v in self._vocab.items()}
        
        super().__init__(
            pad_token=self.PAD_TOKEN,
            bos_token=self.BOS_TOKEN,
            eos_token=self.EOS_TOKEN,
            unk_token=self.UNK_TOKEN,
            **kwargs,
        )

    def _create_default_vocab(self) -> Dict[str, int]:
        special = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]
        return {t: i for i, t in enumerate(special)}

    @property
    def vocab_size(self) -> int:
        return len(self._vocab)

    def get_vocab(self) -> Dict[str, int]:
        return dict(self._vocab)

    def _tokenize(self, text: str) -> List[str]:
        tokens = []
        raw_moves = text.strip().split()
        
        for move in raw_moves:
            if len(move) >= 6:
                # 1. Piece (WP)
                tokens.append(move[:2])
                # 2. From (e2)
                tokens.append(move[2:4])
                # 3. To (e4 or e4(x)) - grab the rest
                tokens.append(move[4:])
            else:
                tokens.append(self.UNK_TOKEN)
        return tokens

    def _convert_token_to_id(self, token: str) -> int:
        return self._vocab.get(token, self._vocab.get(self.UNK_TOKEN))

    def _convert_id_to_token(self, index: int) -> str:
        return self._ids_to_tokens.get(index, self.UNK_TOKEN)

    def convert_tokens_to_string(self, tokens: List[str]) -> str:
        # Filter specials
        filtered = [t for t in tokens if t not in [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]]
        # Join with space. Result: "WP e2 e4 BN g8 f6"
        # The evaluator regex will handle this fine.
        return " ".join(filtered)

    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> tuple:
        if not os.path.isdir(save_directory):
            os.makedirs(save_directory, exist_ok=True)
        vocab_file = os.path.join(save_directory, (filename_prefix + "-" if filename_prefix else "") + "vocab.json")
        with open(vocab_file, "w", encoding="utf-8") as f:
            json.dump(self._vocab, f, ensure_ascii=False, indent=2)
        return (vocab_file,)

    @classmethod
    def build_vocab_from_dataset(cls, dataset_name="dlouapre/lichess_2025-01_1M", split="train", max_samples=10000):
        from datasets import load_dataset
        dataset = load_dataset(dataset_name, split=split, streaming=True)
        
        unique_tokens = set()
        
        print("Building vocabulary...")
        count = 0
        for example in dataset:
            moves = example["text"].split()
            for move in moves:
                if len(move) >= 6:
                    unique_tokens.add(move[:2])    # Piece
                    unique_tokens.add(move[2:4])   # From
                    unique_tokens.add(move[4:])    # To (includes suffix like (x))
            count += 1
            if count >= max_samples:
                break

        special = [cls.PAD_TOKEN, cls.BOS_TOKEN, cls.EOS_TOKEN, cls.UNK_TOKEN]
        # Sort tokens to ensure deterministic IDs
        all_tokens = special + sorted(list(unique_tokens))
        
        vocab = {token: idx for idx, token in enumerate(all_tokens)}
        return cls(vocab=vocab)