""" Chess Move Tokenizer - Component-based approach. This tokenizer decomposes chess moves into atomic components for efficient representation. Each move is broken down into: color, piece type, source square, destination square, and optional annotations (capture, check, promotion, etc.). The vocabulary is built from atomic components rather than full moves, which allows for better generalization and a smaller vocabulary size. """ from __future__ import annotations import json import os from pathlib import Path from typing import Dict, List, Optional import re from transformers import PreTrainedTokenizer # Regular expression to parse extended UCI move notation # Format: [W|B][Piece][from_square][to_square][optional_suffixes] MOVE_PATTERN = re.compile( r"^(?P[WB])" r"(?P[PNBRQK])" r"(?P[a-h][1-8])" r"(?P[a-h][1-8])" r"(?P.*)$" ) class ChessTokenizer(PreTrainedTokenizer): """ Component-based chess move tokenizer. Instead of treating each complete move as a single token, this tokenizer breaks down moves into atomic components (color, piece, squares, annotations). This approach results in a much smaller vocabulary while maintaining the ability to represent all possible chess moves. Example usage: >>> tokenizer = ChessTokenizer() >>> tokens = tokenizer._tokenize("WPe2e4 BPe7e5") >>> # Returns: ['[W]', '[P]', '[e2]', '[e4]', '[B]', '[P]', '[e7]', '[e5]'] """ 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. """ # Set up special token strings self._pad_token = self.PAD_TOKEN self._bos_token = self.BOS_TOKEN self._eos_token = self.EOS_TOKEN self._unk_token = self.UNK_TOKEN # Clean kwargs to prevent conflicts with special tokens # This avoids errors when loading saved tokenizers for token_key in ["pad_token", "bos_token", "eos_token", "unk_token"]: kwargs.pop(token_key, None) # Initialize vocabulary from provided source or create default 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: # Fallback: create minimal vocabulary with component tokens self._vocab = self._create_default_vocab() # Build reverse lookup: token_id -> token_string 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]: """ Construct the default component-based vocabulary. Creates a vocabulary from atomic chess move components: - Special tokens (padding, start, end, unknown) - Color indicators (White/Black) - Piece types (Pawn, Knight, Bishop, Rook, Queen, King) - Board squares (64 squares: a1-h8) - Move annotations (capture, check, checkmate, castling, promotions) """ # Core special tokens special_tokens = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN] # Player color indicators color_tokens = ["[W]", "[B]"] # Chess piece types (note: Bishop uses [BISHOP] to avoid conflict with [B]) piece_tokens = ["[P]", "[N]", "[BISHOP]", "[R]", "[Q]", "[K]"] # All 64 chess board squares square_tokens = [f"[{file}{rank}]" for rank in "12345678" for file in "abcdefgh"] # Move annotations: capture, check, checkmate, castling, promotions annotation_tokens = ["[x]", "[+]", "[#]", "[O-O]", "[O-O-O]", "[prom_Q]", "[prom_R]", "[prom_B]", "[prom_N]"] # Combine all components into vocabulary all_tokens = special_tokens + color_tokens + piece_tokens + square_tokens + annotation_tokens vocab = {token: idx for idx, token in enumerate(all_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. """ return cls() @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. """ return cls() @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]: """ Decompose chess moves into component tokens. Parses each move and breaks it down into atomic components: color, piece, source square, destination square, and annotations. Args: text: Space-separated sequence of moves in extended UCI format. Returns: List of component tokens representing the moves. """ token_list: List[str] = [] move_sequence = text.strip().split() for move_str in move_sequence: # Handle queenside castling (long castling) if "O-O-O" in move_str: player_color = "[W]" if move_str.startswith("W") else "[B]" token_list.append(player_color) token_list.append("[O-O-O]") continue # Handle kingside castling (short castling) if "O-O" in move_str: player_color = "[W]" if move_str.startswith("W") else "[B]" token_list.append(player_color) token_list.append("[O-O]") continue # Parse standard moves using regex match = MOVE_PATTERN.match(move_str) if not match: token_list.append(self.UNK_TOKEN) continue # Extract move components player_color = "[W]" if match.group("side") == "W" else "[B]" piece_type = match.group("piece") from_square = match.group("src") to_square = match.group("dst") move_annotations = match.group("suffix") or "" # Add color and piece token_list.append(player_color) # Handle Bishop separately (B conflicts with Black) if piece_type == "B": token_list.append("[BISHOP]") else: token_list.append(f"[{piece_type}]") # Add squares token_list.append(f"[{from_square}]") token_list.append(f"[{to_square}]") # Process annotations if "x" in move_annotations: token_list.append("[x]") # Capture # Check/checkmate (checkmate takes priority) if "*" in move_annotations: token_list.append("[#]") # Checkmate elif "+" in move_annotations: token_list.append("[+]") # Check # Promotion if "=" in move_annotations: promo_idx = move_annotations.find("=") if promo_idx != -1 and promo_idx + 1 < len(move_annotations): promoted_piece = move_annotations[promo_idx + 1].upper() if promoted_piece in ("Q", "R", "B", "N"): token_list.append(f"[prom_{promoted_piece}]") return token_list def _convert_token_to_id(self, token: str) -> int: """Map token string to its vocabulary ID.""" return self._vocab.get(token, self._vocab.get(self.UNK_TOKEN, 0)) def _convert_id_to_token(self, index: int) -> str: """Map vocabulary ID back to token string.""" return self._ids_to_tokens.get(index, self.UNK_TOKEN) def convert_tokens_to_string(self, tokens: List[str]) -> str: """Reconstruct string from token list, filtering special tokens.""" special_token_set = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN} return " ".join(t for t in tokens if t not in special_token_set) 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]: """ Analyze token frequency distribution in the dataset. Useful for understanding which components appear most frequently and for vocabulary size planning. Args: dataset_name: HuggingFace dataset identifier. split: Which dataset split to analyze. column: Column name containing the game sequences. max_samples: Limit number of samples for faster analysis. Returns: Frequency dictionary: token -> count. """ from collections import Counter from datasets import load_dataset # Load dataset dataset = load_dataset(dataset_name, split=split) # Limit samples if requested if max_samples is not None: dataset = dataset.select(range(min(max_samples, len(dataset)))) # Count component frequencies tokenizer = ChessTokenizer() frequency_counter = Counter() for sample in dataset: component_tokens = tokenizer._tokenize(sample[column]) frequency_counter.update(component_tokens) return dict(frequency_counter)