LLM-course/simple_tokenizer_retrained2

1"""2Custom Chess Tokenizer for the Chess Challenge.3 4This tokenizer treats each move as a single token using the extended UCI notation5from the Lichess dataset (e.g., WPe2e4, BNg8f6).6 7The dataset format uses:8- W/B prefix for White/Black9- Piece letter: P=Pawn, N=Knight, B=Bishop, R=Rook, Q=Queen, K=King10- Source and destination squares (e.g., e2e4)11- Special suffixes: (x)=capture, (+)=check, (+*)=checkmate, (o)/(O)=castling12"""13 14from __future__ import annotations15 16import json17import re18import os19from pathlib import Path20from typing import Dict, List, Optional21 22from transformers import PreTrainedTokenizer23 24 25class ChessTokenizer(PreTrainedTokenizer):26    """27    A custom tokenizer for chess moves using extended UCI notation.28    29    This tokenizer maps each possible chess move to a unique token ID.30    The vocabulary is built from the training dataset to ensure all moves31    encountered during training have a corresponding token.32    33    Example:34        >>> tokenizer = ChessTokenizer()35        >>> tokenizer.encode("WPe2e4 BPe7e5")36        [1, 42, 87, 2]  # [BOS, e2e4, e7e5, EOS]37    """38    39    model_input_names = ["input_ids", "attention_mask"]40    vocab_files_names = {"vocab_file": "vocab.json"}41    42    # Special tokens43    PAD_TOKEN = "[PAD]"44    BOS_TOKEN = "[BOS]"45    EOS_TOKEN = "[EOS]"46    UNK_TOKEN = "[UNK]"47    48    def __init__(49        self,50        vocab_file: Optional[str] = None,51        vocab: Optional[Dict[str, int]] = None,52        **kwargs,53    ):54        """55        Initialize the chess tokenizer.56        57        Args:58            vocab_file: Path to a JSON file containing the vocabulary mapping.59            vocab: Dictionary mapping tokens to IDs (alternative to vocab_file).60            **kwargs: Additional arguments passed to PreTrainedTokenizer.61        """62        # Initialize special tokens63        self._pad_token = self.PAD_TOKEN64        self._bos_token = self.BOS_TOKEN65        self._eos_token = self.EOS_TOKEN66        self._unk_token = self.UNK_TOKEN67 68        # Remove any duplicate special-token entries passed through kwargs69        # to avoid "multiple values for keyword" errors when loading from disk.70        kwargs.pop("pad_token", None)71        kwargs.pop("bos_token", None)72        kwargs.pop("eos_token", None)73        kwargs.pop("unk_token", None)74        75        # Load or create vocabulary76        if vocab is not None:77            self._vocab = vocab78        elif vocab_file is not None and os.path.exists(vocab_file):79            with open(vocab_file, "r", encoding="utf-8") as f:80                self._vocab = json.load(f)81        else:82            # Create a minimal vocabulary with just special tokens83            # The full vocabulary should be built from the dataset84            self._vocab = self._create_default_vocab()85        86        # Create reverse mapping87        self._ids_to_tokens = {v: k for k, v in self._vocab.items()}88        89        # Call parent init AFTER setting up vocab90        super().__init__(91            pad_token=self._pad_token,92            bos_token=self._bos_token,93            eos_token=self._eos_token,94            unk_token=self._unk_token,95            **kwargs,96        )97    98    def _create_default_vocab(self) -> Dict[str, int]:99        """100        Create a minimal default vocabulary with just special tokens.101        102        For the full vocabulary, use `build_vocab_from_dataset()`.103        This minimal vocab is just a placeholder - you should build from data.104        """105        special_tokens = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]106        vocab = {token: idx for idx, token in enumerate(special_tokens)}107        return vocab108    109    @classmethod110    def build_vocab_from_iterator(111        cls,112        iterator,113        min_frequency: int = 1,114    ) -> "ChessTokenizer":115        """116        Build a tokenizer vocabulary from an iterator of game strings.117        118        Args:119            iterator: An iterator yielding game strings (space-separated moves).120            min_frequency: Minimum frequency for a token to be included.121        122        Returns:123            A ChessTokenizer with the built vocabulary.124        """125 126 127        # from collections import Counter128        #129        # token_counts = Counter()130        131        # for game in iterator:132        #     moves = game.strip().split()133        #     token_counts.update(moves)134        #135        #136        # # Filter by frequency137        # tokens = [138        #     token for token, count in token_counts.items()139        #     if count >= min_frequency140        # ]141        #142        # # Sort for reproducibility143        # tokens = sorted(tokens)144        145        # Build vocabulary146        special_tokens = [cls.PAD_TOKEN, cls.BOS_TOKEN, cls.EOS_TOKEN, cls.UNK_TOKEN]147        piece = ['K', 'Q', 'R', 'B', 'N', 'P']148        move = ['a1', 'a2', 'a3', 'a4', 'a5', 'a6', 'a7', 'a8', 'b1', 'b2', 'b3', 'b4', 'b5', 'b6', 'b7', 'b8', 'c1', 'c2', 'c3', 'c4', 'c5', 'c6', 'c7', 'c8', 'd1', 'd2', 'd3', 'd4', 'd5', 'd6', 'd7', 'd8', 'e1', 'e2', 'e3', 'e4', 'e5', 'e6', 'e7', 'e8', 'f1', 'f2', 'f3', 'f4', 'f5', 'f6', 'f7', 'f8', 'g1', 'g2', 'g3', 'g4', 'g5', 'g6', 'g7', 'g8', 'h1', 'h2', 'h3', 'h4', 'h5', 'h6', 'h7', 'h8']149 150        # vocab = {token: idx for idx, token in enumerate(special_tokens + tokens)}151        vocab = {token: idx for idx, token in enumerate(special_tokens + piece + move)}152        153        return cls(vocab=vocab)154    155    @classmethod156    def build_vocab_from_dataset(157        cls,158        dataset_name: str = "dlouapre/lichess_2025-01_1M",159        split: str = "train",160        column: str = "text",161        min_frequency: int = 500,162        max_samples: Optional[int] = 100000,163    ) -> "ChessTokenizer":164        """165        Build a tokenizer vocabulary from a Hugging Face dataset.166        167        Args:168            dataset_name: Name of the dataset on Hugging Face Hub.169            split: Dataset split to use.170            column: Column containing the game strings.171            min_frequency: Minimum frequency for a token to be included (default: 500).172            max_samples: Maximum number of samples to process (default: 100k).173        174        Returns:175            A ChessTokenizer with the built vocabulary.176        """177        from datasets import load_dataset178        179        dataset = load_dataset(dataset_name, split=split)180        181        if max_samples is not None:182            dataset = dataset.select(range(min(max_samples, len(dataset))))183        184        def game_iterator():185            for example in dataset:186                yield example[column]187        188        return cls.build_vocab_from_iterator(game_iterator(), min_frequency=min_frequency)189    190    @property191    def vocab_size(self) -> int:192        """Return the size of the vocabulary."""193        return len(self._vocab)194    195    def get_vocab(self) -> Dict[str, int]:196        """Return the vocabulary as a dictionary."""197        return dict(self._vocab)198    199    def _tokenize(self, text: str) -> List[str]:200        """201        Tokenize a string of moves into a list of tokens.202        203        Args:204            text: A string of space-separated moves.205        206        Returns:207            List of move tokens.208        """209 210        regex = r"^[WB]([KQRBNP])([a-h][1-8])([a-h][1-8])"211 212        tokens = []213        for move in text.strip().split():214            match = re.search(regex, move)215            if match:216                tokens += list(match.groups())217            else:218                tokens += self.UNK_TOKEN219 220        return tokens221    222    def _convert_token_to_id(self, token: str) -> int:223        """Convert a token to its ID."""224        return self._vocab.get(token, self._vocab.get(self.UNK_TOKEN, 0))225    226    def _convert_id_to_token(self, index: int) -> str:227        """Convert an ID to its token."""228        return self._ids_to_tokens.get(index, self.UNK_TOKEN)229    230    def convert_tokens_to_string(self, tokens: List[str]) -> str:231        """Convert a list of tokens back to a string."""232        # Filter out special tokens for cleaner output233        special = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}234        return " ".join(t for t in tokens if t not in special)235    236    def save_vocabulary(237        self,238        save_directory: str,239        filename_prefix: Optional[str] = None,240    ) -> tuple:241        """242        Save the vocabulary to a JSON file.243        244        Args:245            save_directory: Directory to save the vocabulary.246            filename_prefix: Optional prefix for the filename.247        248        Returns:249            Tuple containing the path to the saved vocabulary file.250        """251        if not os.path.isdir(save_directory):252            os.makedirs(save_directory, exist_ok=True)253        254        vocab_file = os.path.join(255            save_directory,256            (filename_prefix + "-" if filename_prefix else "") + "vocab.json",257        )258        259        with open(vocab_file, "w", encoding="utf-8") as f:260            json.dump(self._vocab, f, ensure_ascii=False, indent=2)261        262        return (vocab_file,)263 264 265def count_vocab_from_dataset(266    dataset_name: str = "dlouapre/lichess_2025-01_1M",267    split: str = "train",268    column: str = "text",269    max_samples: Optional[int] = 10000,270) -> Dict[str, int]:271    """272    Count token frequencies in a dataset (useful for vocabulary analysis).273    274    Args:275        dataset_name: Name of the dataset on Hugging Face Hub.276        split: Dataset split to use.277        column: Column containing the game strings.278        max_samples: Maximum number of samples to process.279    280    Returns:281        Dictionary mapping tokens to their frequencies.282    """283    from collections import Counter284    from datasets import load_dataset285    286    dataset = load_dataset(dataset_name, split=split)287    288    if max_samples is not None:289        dataset = dataset.select(range(min(max_samples, len(dataset))))290    291    token_counts = Counter()292    293    for example in dataset:294        moves = example[column].strip().split()295        token_counts.update(moves)296    297    return dict(token_counts)298