tokenizer.py

"""
Custom Chess Tokenizer for the Chess Challenge.

Move-level tokenizer:
- Each move string is ONE token, e.g. "WPe2e4", "BNg8f6", "WBb5c6(x)".

Key improvement vs baseline:
- Adds `max_vocab_size` to cap vocabulary size (very important for <1M params).
- Keeps the TOP-K most frequent moves (after min_frequency filter).
- Registers for AutoTokenizer so server-side loading works.
"""

from __future__ import annotations

import json
import os
from typing import Dict, List, Optional, Tuple

from transformers import PreTrainedTokenizer


class ChessTokenizer(PreTrainedTokenizer):
    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,
    ):
        # Set special tokens
        self._pad_token = self.PAD_TOKEN
        self._bos_token = self.BOS_TOKEN
        self._eos_token = self.EOS_TOKEN
        self._unk_token = self.UNK_TOKEN

        # Remove duplicates in kwargs (important when loading)
        kwargs.pop("pad_token", None)
        kwargs.pop("bos_token", None)
        kwargs.pop("eos_token", None)
        kwargs.pop("unk_token", None)

        # 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_tokens = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]
        return {tok: i for i, tok in enumerate(special_tokens)}

    @classmethod
    def build_vocab_from_iterator(
        cls,
        iterator,
        min_frequency: int = 1,
        max_vocab_size: int = 5000,
    ) -> "ChessTokenizer":
        """
        Build vocabulary from an iterator of game strings.

        Strategy:
        - Count move frequency
        - Filter by min_frequency
        - Take TOP-K most frequent moves (K = max_vocab_size - #special_tokens)

        This avoids vocab explosion (which breaks the <1M parameter constraint).
        """
        from collections import Counter

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

        # Filter by frequency
        items = [(tok, c) for tok, c in token_counts.items() if c >= min_frequency]

        # Sort by frequency desc then token for reproducibility
        items.sort(key=lambda x: (-x[1], x[0]))

        special_tokens = [cls.PAD_TOKEN, cls.BOS_TOKEN, cls.EOS_TOKEN, cls.UNK_TOKEN]
        budget = max_vocab_size - len(special_tokens)
        if budget <= 0:
            raise ValueError("max_vocab_size must be > number of special tokens")

        top_tokens = [tok for tok, _ in items[:budget]]

        vocab = {tok: i for i, tok in enumerate(special_tokens + top_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 = 500,
        max_samples: Optional[int] = 100000,
        max_vocab_size: int = 5000,
    ) -> "ChessTokenizer":
        """
        Build vocabulary from a HF dataset.

        IMPORTANT:
        - `max_vocab_size` caps final vocab size (including special tokens).
        - `min_frequency` filters extremely rare moves before top-k selection.
        """
        from datasets import load_dataset

        dataset = load_dataset(dataset_name, split=split)

        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,
            max_vocab_size=max_vocab_size,
        )

    @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]:
        return text.strip().split()

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


# IMPORTANT for server-side loading on HF
ChessTokenizer.register_for_auto_class("AutoTokenizer")