tokenizer.py

"""
Custom Chess Tokenizer for the Chess Challenge with TRM optimizations.
This tokenizer treats each move as a single token using the extended UCI notation.
Optimized for smaller vocabulary size to work better with TRM's tiny networks.
"""
from __future__ import annotations
import json
import os
from typing import Dict, List, Optional
from transformers import PreTrainedTokenizer


class ChessTokenizer(PreTrainedTokenizer):
    """
    A custom tokenizer for chess moves using extended UCI notation.
    Optimized for TRM with aggressive vocabulary pruning.
    """
    model_input_names = ["input_ids", "attention_mask"]
    vocab_files_names = {"vocab_file": "vocab.json"}

    # Special tokens
    PAD_TOKEN = "[PAD]"
    BOS_TOKEN = "[BOS]"
    EOS_TOKEN = "[EOS]"
    UNK_TOKEN = "[UNK]"

    def __init__(
        self,
        vocab_file: Optional[str] = None,
        vocab: Optional[Dict[str, int]] = None,
        **kwargs,
    ):
        """Initialize the chess tokenizer."""
        # Remove special token kwargs if they exist (to avoid duplicates)
        kwargs.pop('pad_token', None)
        kwargs.pop('bos_token', None)
        kwargs.pop('eos_token', None)
        kwargs.pop('unk_token', None)
        
        self._pad_token = self.PAD_TOKEN
        self._bos_token = self.BOS_TOKEN
        self._eos_token = self.EOS_TOKEN
        self._unk_token = self.UNK_TOKEN

        # Load or create vocabulary
        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()

        # Create reverse mapping
        self._ids_to_tokens = {v: k for k, v in self._vocab.items()}

        # Call parent init AFTER setting up vocab
        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]:
        """Create a minimal default vocabulary with just special tokens."""
        special_tokens = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]
        vocab = {token: idx for idx, token in enumerate(special_tokens)}
        return vocab

    @classmethod
    def build_vocab_from_iterator(
        cls, iterator, min_frequency: int = 1
    ) -> "ChessTokenizer":
        """Build a tokenizer vocabulary from an iterator of game strings."""
        from collections import Counter

        token_counts = Counter()
        for game in iterator:
            moves = game.strip().split()
            token_counts.update(moves)

        # Filter by frequency
        tokens = [
            token for token, count in token_counts.items() if count >= min_frequency
        ]

        # Sort for reproducibility
        tokens = sorted(tokens)

        # Build vocabulary
        special_tokens = [cls.PAD_TOKEN, cls.BOS_TOKEN, cls.EOS_TOKEN, cls.UNK_TOKEN]
        vocab = {token: idx for idx, token in enumerate(special_tokens + tokens)}

        return cls(vocab=vocab)

    @classmethod
    def build_vocab_from_dataset(
        cls,
        dataset_name: str = "dlouapre/lichess_2025-01_1M",
        split: str = "train",
        column: str = "text",
        min_frequency: int = 1000,  # Increased for TRM (smaller vocab)
        max_samples: Optional[int] = 200000,  # More samples for better coverage
    ) -> "ChessTokenizer":
        """
        Build tokenizer vocabulary optimized for TRM.
        Uses higher min_frequency to create a smaller, more focused vocabulary.
        """
        from datasets import load_dataset

        dataset = load_dataset(dataset_name, split=split, cache_dir=os.environ["HF_DATASETS_CACHE"])
        if max_samples is not None:
            dataset = dataset.select(range(min(max_samples, len(dataset))))

        def game_iterator():
            for example in dataset:
                yield example[column]

        return cls.build_vocab_from_iterator(
            game_iterator(), min_frequency=min_frequency
        )

    @property
    def vocab_size(self) -> int:
        """Return the size of the vocabulary."""
        return len(self._vocab)

    def get_vocab(self) -> Dict[str, int]:
        """Return the vocabulary as a dictionary."""
        return dict(self._vocab)

    def _tokenize(self, text: str) -> List[str]:
        """Tokenize a string of moves into a list of tokens."""
        return text.strip().split()

    def _convert_token_to_id(self, token: str) -> int:
        """Convert a token to its ID."""
        return self._vocab.get(token, self._vocab.get(self.UNK_TOKEN, 0))

    def _convert_id_to_token(self, index: int) -> str:
        """Convert an ID to its token."""
        return self._ids_to_tokens.get(index, self.UNK_TOKEN)

    def convert_tokens_to_string(self, tokens: List[str]) -> str:
        """Convert a list of tokens back to a string."""
        special = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}
        return " ".join(t for t in tokens if t not in special)

    def save_vocabulary(
        self, save_directory: str, filename_prefix: Optional[str] = None
    ) -> tuple:
        """Save the vocabulary to a JSON file."""
        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,)