""" Custom Chess Tokenizer for the Chess Challenge. This tokenizer uses a compact hybrid scheme optimized for small models: - Frequent moves are single tokens (e.g., WPe2e4). - Rare moves fall back to two tokens: piece+from (e.g., WPe2) and to-square (e.g., e4). - Promotions add a third token (q/r/b/n). The dataset format uses: - W/B prefix for White/Black - Piece letter: P=Pawn, N=Knight, B=Bishop, R=Rook, Q=Queen, K=King - Source and destination squares (e.g., e2e4) - Special suffixes: (x)=capture, (+)=check, (+*)=checkmate, (o)/(O)=castling """ from __future__ import annotations import json import os import re from pathlib import Path from typing import Dict, List, Optional, Tuple from transformers import PreTrainedTokenizer class ChessTokenizer(PreTrainedTokenizer): """ A custom tokenizer for chess moves using extended UCI notation. This tokenizer uses a compact base vocabulary (piece+from, to-square, promotion tokens) and optionally adds frequent full-move tokens for shorter sequences and better sample efficiency. Example: >>> tokenizer = ChessTokenizer() >>> tokenizer.encode("WPe2e4 BPe7e5") [1, 42, 87, 2] # [BOS, e2e4, e7e5, EOS] """ 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]" _MOVE_RE = re.compile( r"^(?P[WB])(?P[PNBRQK])(?P[a-h][1-8])(?P[a-h][1-8])(?P.*)$" ) _PROMO_RE = re.compile(r"=([NBRQnrbq])") def __init__( self, vocab_file: Optional[str] = None, vocab: Optional[Dict[str, int]] = None, **kwargs, ): """ Initialize the chess tokenizer. Args: vocab_file: Path to a JSON file containing the vocabulary mapping. vocab: Dictionary mapping tokens to IDs (alternative to vocab_file). **kwargs: Additional arguments passed to PreTrainedTokenizer. """ # Initialize 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 any duplicate special-token entries passed through kwargs # to avoid "multiple values for keyword" errors when loading from disk. kwargs.pop("pad_token", None) kwargs.pop("bos_token", None) kwargs.pop("eos_token", None) kwargs.pop("unk_token", None) # 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: # Create a compact default vocabulary that can tokenize any move 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 compact default vocabulary with full move coverage. For better compression, use `build_vocab_from_dataset()` to add frequent full-move tokens. """ tokens = self._create_base_vocab_tokens() return {token: idx for idx, token in enumerate(tokens)} @classmethod def _create_base_vocab_tokens(cls) -> List[str]: special_tokens = [cls.PAD_TOKEN, cls.BOS_TOKEN, cls.EOS_TOKEN, cls.UNK_TOKEN] pieces = ["P", "N", "B", "R", "Q", "K"] colors = ["W", "B"] files = "abcdefgh" ranks = "12345678" squares = [f"{f}{r}" for f in files for r in ranks] piece_from_tokens = [f"{c}{p}{sq}" for c in colors for p in pieces for sq in squares] to_tokens = squares promo_tokens = ["q", "r", "b", "n"] return special_tokens + piece_from_tokens + to_tokens + promo_tokens @classmethod def _parse_move(cls, token: str) -> Optional[Tuple[str, str, str, str, Optional[str]]]: match = cls._MOVE_RE.match(token) if not match: return None color = match.group("color") piece = match.group("piece") from_sq = match.group("from") to_sq = match.group("to") rest = match.group("rest") promo_match = cls._PROMO_RE.search(rest) promo = promo_match.group(1).upper() if promo_match else None return color, piece, from_sq, to_sq, promo @classmethod def build_vocab_from_iterator( cls, iterator, min_frequency: int = 1, max_full_move_tokens: Optional[int] = 1200, ) -> "ChessTokenizer": """ Build a tokenizer vocabulary from an iterator of game strings. Args: iterator: An iterator yielding game strings (space-separated moves). min_frequency: Minimum frequency for a token to be included. max_full_move_tokens: Maximum number of full-move tokens to keep. Returns: A ChessTokenizer with the built vocabulary. """ from collections import Counter token_counts = Counter() for game in iterator: moves = game.strip().split() for move in moves: parsed = cls._parse_move(move) if not parsed: continue color, piece, from_sq, to_sq, promo = parsed if promo: continue token_counts[f"{color}{piece}{from_sq}{to_sq}"] += 1 # Filter by frequency tokens = [ token for token, count in token_counts.items() if count >= min_frequency ] # Sort by frequency, then lexicographically for reproducibility tokens.sort(key=lambda t: (-token_counts[t], t)) if max_full_move_tokens is not None: tokens = tokens[:max_full_move_tokens] base_tokens = cls._create_base_vocab_tokens() vocab = {token: idx for idx, token in enumerate(base_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 = 500, max_samples: Optional[int] = 100000, max_full_move_tokens: Optional[int] = 1200, ) -> "ChessTokenizer": """ Build a tokenizer vocabulary from a Hugging Face dataset. Args: dataset_name: Name of the dataset on Hugging Face Hub. split: Dataset split to use. column: Column containing the game strings. min_frequency: Minimum frequency for a token to be included (default: 500). max_samples: Maximum number of samples to process (default: 100k). max_full_move_tokens: Maximum number of full-move tokens to keep. Returns: A ChessTokenizer with the built vocabulary. """ 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_full_move_tokens=max_full_move_tokens, ) @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. Args: text: A string of space-separated moves. Returns: List of move tokens. """ raw = text.strip() if not raw: return [] parts = raw.split() out: List[str] = [] special = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN} for part in parts: if part in special: out.append(part) continue parsed = self._parse_move(part) if not parsed: out.append(self.UNK_TOKEN) continue color, piece, from_sq, to_sq, promo = parsed full_move = f"{color}{piece}{from_sq}{to_sq}" if promo is None and full_move in self._vocab: out.append(full_move) continue piece_from = f"{color}{piece}{from_sq}" to_token = f"{to_sq}" out.append(piece_from if piece_from in self._vocab else self.UNK_TOKEN) out.append(to_token if to_token in self._vocab else self.UNK_TOKEN) if promo: promo_token = promo.lower() out.append(promo_token if promo_token in self._vocab else self.UNK_TOKEN) return out 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.""" # Filter out special tokens for cleaner output 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. Args: save_directory: Directory to save the vocabulary. filename_prefix: Optional prefix for the filename. Returns: Tuple containing the path to the saved vocabulary 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,) def count_vocab_from_dataset( dataset_name: str = "dlouapre/lichess_2025-01_1M", split: str = "train", column: str = "text", max_samples: Optional[int] = 10000, ) -> Dict[str, int]: """ Count normalized move frequencies in a dataset (useful for vocabulary analysis). Args: dataset_name: Name of the dataset on Hugging Face Hub. split: Dataset split to use. column: Column containing the game strings. max_samples: Maximum number of samples to process. Returns: Dictionary mapping normalized full-move tokens to their frequencies. """ from collections import Counter 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)))) token_counts = Counter() for example in dataset: moves = example[column].strip().split() for move in moves: parsed = ChessTokenizer._parse_move(move) if not parsed: continue color, piece, from_sq, to_sq, promo = parsed if promo: continue token_counts[f"{color}{piece}{from_sq}{to_sq}"] += 1 return dict(token_counts)