""" 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 typing import Dict, List, Optional from transformers import PreTrainedTokenizer class ChessTokenizerOld(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 ChessTokenizer(PreTrainedTokenizer): """ A sophisticated chess tokenizer that decomposes moves into components. Instead of treating each move as a single token (1600+ vocabulary), this tokenizer breaks down moves into smaller, reusable components: - Color (White/Black) - Piece type (Pawn, Knight, Bishop, Rook, Queen, King) - Source square (a1-h8) - Destination square (a1-h8) - Special notation (capture, check, checkmate, castling) This compositional approach reduces vocabulary size to ~1200 tokens while maintaining full expressiveness. Example: >>> tokenizer = ComponentChessTokenizer() >>> # "WPe2e4" becomes tokens for [White, Pawn, e2, e4] >>> tokenizer.encode("WPe2e4 BPe7e5") [1, 5, 10, 20, 28, 6, 10, 21, 29, 2] # [BOS, W, P, e2, e4, B, P, e7, e5, 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]" # Component tokens - these are fixed COLOR_TOKENS = ["[W]", "[B]"] # White, Black PIECE_TOKENS = ["[P]", "[N]", "[B]", "[R]", "[Q]", "[K]"] # Pawn, Knight, Bishop, Rook, Queen, King SQUARE_TOKENS = [f"[{file}{rank}]" for file in "abcdefgh" for rank in "12345678"] # 64 squares SPECIAL_TOKENS_MOVE = [ "[x]", # Capture "[+]", # Check "[#+]", # Checkmate "[o]", # Kingside castling (short) "[O]", # Queenside castling (long) ] def __init__( self, vocab_file: Optional[str] = None, vocab: Optional[Dict[str, int]] = None, **kwargs, ): """ Initialize the component-based 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 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: self._vocab = self._create_component_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_component_vocab(self) -> Dict[str, int]: """ Create a vocabulary from pre-defined components. Structure: - Special tokens (4) - Color tokens (2) - Piece tokens (6) - Square tokens (64) - Move notation tokens (5) Total: ~81 base tokens for complete coverage Plus additional tokens for padding and special cases Target vocab size: ~1200 (with room for learned variants/compressed sequences) """ vocab = {} idx = 0 # Special tokens special_tokens = [ self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN, ] for token in special_tokens: vocab[token] = idx idx += 1 # Color tokens for token in self.COLOR_TOKENS: vocab[token] = idx idx += 1 # Piece tokens for token in self.PIECE_TOKENS: vocab[token] = idx idx += 1 # Square tokens for token in self.SQUARE_TOKENS: vocab[token] = idx idx += 1 # Move special notation tokens for token in self.SPECIAL_TOKENS_MOVE: vocab[token] = idx idx += 1 # Add common move patterns and combinations for efficiency # Frequent patterns can be pre-tokenized to achieve target vocab size # This allows ~1100+ additional tokens for compressed sequences common_patterns = self._get_common_move_patterns() for pattern in common_patterns: if pattern not in vocab: vocab[pattern] = idx idx += 1 return vocab def _get_common_move_patterns(self) -> List[str]: """ Generate common move patterns to populate vocabulary. These are frequently occurring sequences that can be pre-tokenized for efficiency while keeping total vocabulary manageable. """ patterns = [] # Common opening moves (e.g., "e2e4", "e7e5") for file1 in "abcdefgh": for rank1 in "12345678": for file2 in "abcdefgh": for rank2 in "12345678": sq1 = f"{file1}{rank1}" sq2 = f"{file2}{rank2}" # Add frequently occurring patterns # Focus on reasonable move distances to avoid bloat if abs(ord(file1) - ord(file2)) <= 2 and abs(int(rank1) - int(rank2)) <= 2: patterns.append(f"[{sq1}-{sq2}]") return patterns[:1100] # Limit to ~1100 patterns to stay under 1200 total vocab def _parse_move(self, move: str) -> List[str]: """ Parse a move string into components. Examples: "WPe2e4" -> ["[W]", "[P]", "[e2]", "[e4]"] "BNg8f6x" -> ["[B]", "[N]", "[g8]", "[f6]", "[x]"] "WKe1g1o" -> ["[W]", "[K]", "[e1]", "[g1]", "[o]"] Args: move: A move string in extended UCI format. Returns: List of component tokens. """ if not move or len(move) < 4: return [self.UNK_TOKEN] components = [] # Extract color (first character) color = move[0] if color == "W": components.append("[W]") elif color == "B": components.append("[B]") else: return [self.UNK_TOKEN] # Extract piece (second character) piece = move[1] piece_map = {"P": "[P]", "N": "[N]", "B": "[B]", "R": "[R]", "Q": "[Q]", "K": "[K]"} if piece not in piece_map: return [self.UNK_TOKEN] components.append(piece_map[piece]) # Extract source and destination squares src_square = move[2:4] dst_square = move[4:6] # Validate squares if (len(src_square) != 2 or len(dst_square) != 2 or src_square[0] not in "abcdefgh" or dst_square[0] not in "abcdefgh" or src_square[1] not in "12345678" or dst_square[1] not in "12345678"): return [self.UNK_TOKEN] components.append(f"[{src_square}]") components.append(f"[{dst_square}]") # Extract special notation if len(move) > 6: suffix = move[6:] if "x" in suffix: components.append("[x]") if "+*" in suffix: components.append("[#+]") elif "+" in suffix: components.append("[+]") if "o" in suffix.lower(): if "O" in move: components.append("[O]") # Queenside castling else: components.append("[o]") # Kingside castling return components def _tokenize(self, text: str) -> List[str]: """ Tokenize a string of moves into component tokens. Args: text: A string of space-separated moves. Returns: List of component tokens. """ moves = text.strip().split() tokens = [] for move in moves: components = self._parse_move(move) tokens.extend(components) return tokens 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 representation.""" # Filter out special tokens and brackets for cleaner output cleaned = [] for t in tokens: if t not in {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}: # Remove brackets if present t = t.strip("[]") if t: cleaned.append(t) return " ".join(cleaned) 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,) @classmethod def build_vocab_from_iterator( cls, iterator, min_frequency: int = 1, ) -> "ComponentChessTokenizer": """ Build a tokenizer vocabulary from an iterator of game strings. This method decomposes moves into components and builds the vocabulary from the component tokens. Args: iterator: An iterator yielding game strings (space-separated moves). min_frequency: Minimum frequency for a component token to be included. Returns: A ComponentChessTokenizer with the built vocabulary. """ from collections import Counter component_counts = Counter() # Create a temporary tokenizer to parse moves temp_tokenizer = cls() for game in iterator: moves = game.strip().split() for move in moves: components = temp_tokenizer._parse_move(move) component_counts.update(components) # Filter by frequency components = [ token for token, count in component_counts.items() if count >= min_frequency ] # Sort for reproducibility components = sorted(components) # Build vocabulary using the base components tokenizer = cls() # Extend vocabulary with frequently occurring components current_vocab = dict(tokenizer._vocab) idx = len(current_vocab) for component in components: if component not in current_vocab: current_vocab[component] = idx idx += 1 return cls(vocab=current_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, ) -> "ComponentChessTokenizer": """ Build a tokenizer vocabulary from a Hugging Face dataset. This method decomposes moves into components and builds the vocabulary from the component tokens found in the 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 component token to be included (default: 500). max_samples: Maximum number of samples to process (default: 100k). Returns: A ComponentChessTokenizer 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 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)