""" Custom Chess Tokenizer for the Chess Challenge. This tokenizer treats each move as a single token using the extended UCI notation from the Lichess dataset (e.g., WPe2e4, BNg8f6). 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 from pathlib import Path from token import OP from typing import Dict, List, Optional from transformers import PreTrainedTokenizer import re class ChessTokenizer(PreTrainedTokenizer): """ A custom tokenizer for chess moves using extended UCI notation. This tokenizer maps each possible chess move to a unique token ID. The vocabulary is built from the training dataset to ensure all moves encountered during training have a corresponding token. 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]" 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 minimal vocabulary with just special tokens # The full vocabulary should be built from the dataset 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. For the full vocabulary, use `build_vocab_from_dataset()`. This minimal vocab is just a placeholder - you should build from data. """ 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. Args: iterator: An iterator yielding game strings (space-separated moves). min_frequency: Minimum frequency for a token to be included. Returns: A ChessTokenizer with the built vocabulary. """ 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 = 500, max_samples: Optional[int] = 100000, ) -> "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). 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) @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. """ 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.""" # 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 token 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 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() token_counts.update(moves) return dict(token_counts) class CoordinateTokenizer(ChessTokenizer): def __init__(self, **kwargs): squares = [f"{f}{r}" for f in "abcdefgh" for r in "12345678"] promotions = ["q", "r", "b", "n"] control = ["[PAD]", "[BOS]", "[EOS]", "[UNK]"] vocab_list = control + squares + promotions self._vocab = {t: i for i, t in enumerate(vocab_list)} self._ids_to_token = {i: t for t, i in self._vocab.items()} super().__init__( vocab=self._vocab, pad_token="[PAD]", bos_token="[BOS]", eos_token="[EOS]", unk_token="[UNK]", truncation_side="left", **kwargs ) def _tokenize(self, text: str) -> List[str]: raw_moves = text.strip().split() tokens = [] for raw_move in raw_moves: squares = re.findall(r'[a-h][1-8]', raw_move) tokens.extend(squares) if "=" in raw_move: idx = raw_move.index("=") if idx + 1 < len(raw_move): tokens.append(raw_move[idx+1].lower()) elif "q" in raw_move[-2:].lower(): tokens.append(raw_move[-1].lower()) return tokens class CoordinateChessTokenizer(PreTrainedTokenizer): """ Tokenizer that decomposes chess moves into coordinate components. Example: WPe2e4 -> ['e2', 'e4'] WPa7a8q -> ['a7', 'a8', 'q'] # pawn promotion Vocabulary size: 72 tokens - 64 squares (a1-h8) - 4 promotions (q, r, b, n) - 4 special tokens """ model_input_names = ["input_ids", "attention_mask"] vocab_files_names = {"vocab_file": "vocab.json"} PAD_TOKEN = "[PAD]" BOS_TOKEN = "[BOS]" EOS_TOKEN = "[EOS]" UNK_TOKEN = "[UNK]" # Regex to extract from-square, to-square, and optional promotion MOVE_PATTERN = re.compile(r'([a-h][1-8])([a-h][1-8])([qrbn])?') def __init__(self, vocab_file: Optional[str] = None, **kwargs): # Remove duplicate special token kwargs kwargs.pop("pad_token", None) kwargs.pop("bos_token", None) kwargs.pop("eos_token", None) kwargs.pop("unk_token", None) # Build fixed vocabulary if 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_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_vocab(self) -> Dict[str, int]: """Create fixed vocabulary of 72 tokens.""" tokens = [ self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN, ] # Add all 64 squares for file in 'abcdefgh': for rank in '12345678': tokens.append(f"{file}{rank}") # Add promotion pieces tokens.extend(['q', 'r', 'b', 'n']) return {token: idx for idx, token in enumerate(tokens)} @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]: """ Tokenize move string into coordinate components. Args: text: Space-separated moves like "WPe2e4 BNg8f6" Returns: List of coordinate tokens: ['e2', 'e4', 'g8', 'f6'] """ tokens = [] raw_moves = text.strip().split() for move in raw_moves: match = self.MOVE_PATTERN.search(move) if match: from_sq, to_sq, promotion = match.groups() tokens.append(from_sq) tokens.append(to_sq) if promotion: tokens.append(promotion) return tokens 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: """Reconstruct moves from coordinate tokens.""" special = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN} clean = [t for t in tokens if t not in special] # Group into moves (2 or 3 tokens per move) moves = [] i = 0 while i < len(clean): if i + 1 < len(clean): move = clean[i] + clean[i + 1] i += 2 # Check for promotion if i < len(clean) and clean[i] in ['q', 'r', 'b', 'n']: move += clean[i] i += 1 moves.append(move) else: i += 1 return " ".join(moves) 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,) class EnhancedCoordinateTokenizer(CoordinateChessTokenizer): """ Extended version that preserves piece information as optional metadata. Vocabulary: 76 tokens (adds W, B, P, N, B, R, Q, K but makes them optional) Use this if you want to preserve color/piece info with minimal vocab growth. """ def _create_vocab(self) -> Dict[str, int]: vocab = super()._create_vocab() # Add optional color and piece tokens piece_tokens = ['W', 'B', 'P', 'N', 'R', 'Q', 'K'] # Note: B appears in both contexts next_id = len(vocab) for token in piece_tokens: if token not in vocab: vocab[token] = next_id next_id += 1 return vocab def _tokenize(self, text: str) -> List[str]: """ Optionally include piece info: WPe2e4 -> ['W', 'P', 'e2', 'e4'] Or strip it for minimal version: WPe2e4 -> ['e2', 'e4'] """ tokens = [] raw_moves = text.strip().split() for move in raw_moves: # Extract color and piece if present if len(move) >= 2 and move[0] in 'WB' and move[1] in 'PNBRQK': # Uncomment to include piece info (increases sequence length): # tokens.extend([move[0], move[1]]) pass # Extract coordinates match = self.MOVE_PATTERN.search(move) if match: from_sq, to_sq, promotion = match.groups() tokens.append(from_sq) tokens.append(to_sq) if promotion: tokens.append(promotion) return tokens class SanitizedChessTokenizer(ChessTokenizer): # Strategy: # 1. Strip suffixes: (, ), x, +, *, o, O, E # 2. Strip prefixes: W or B followed by P, N, B, R, Q, K # Regex: ^[WB][PNBRQK] matches the start of the string # We can use a single regex to find the "Pure Move" part. # We look for the square-to-square pattern (e.g., e2e4) and optional promotion (q,r,b,n) # This is safer than stripping because it ignores all noise around the move. MOVE_PATTERN = re.compile(r'([a-h][1-8][a-h][1-8][qrbn]?)') def _sanitize(self, text: str) -> str: # Extract just the move part (e.g., "WPe2e4(x)" -> "e2e4") match = self.MOVE_PATTERN.search(text) if match: return match.group(1) return self.unk_token # Fallback if no valid move found def _tokenize(self, text: str) -> List[str]: # Tokenize by splitting space, then extracting the move tokens = [] for t in text.strip().split(): clean = self._sanitize(t) if clean != self.unk_token: tokens.append(clean) return tokens @classmethod def build_vocab_from_iterator(cls, iterator, min_frequency: int = 1) -> "SanitizedChessTokenizer": from collections import Counter token_counts = Counter() for game in iterator: moves = game.strip().split() # Extract only the Pure UCI part clean_moves = [] for m in moves: match = cls.MOVE_PATTERN.search(m) if match: clean_moves.append(match.group(1)) token_counts.update(clean_moves) # Filter by frequency tokens = [ token for token, count in token_counts.items() if count >= min_frequency ] 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)