""" Custom Chess Tokenizer for the Chess Challenge (Improved Version). This tokenizer uses an atomic approach: each move is decomposed into component tokens: - Side: [W] or [B] - Piece: [P], [N], [B], [R], [Q], [K] - Squares: [a1] through [h8] (64 tokens) - Flags: [x] (capture), [+] (check), [#] (mate), [O-O], [O-O-O] (castling) - Promotions: [=q], [=r], [=b], [=n] This approach reduces vocabulary from ~1200 to 84 tokens, saving ~142K parameters! Example: "WPe2e4" -> ["[W]", "[P]", "[e2]", "[e4]"] "BNg8f6(x)" -> ["[B]", "[N]", "[g8]", "[f6]", "[x]"] """ 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 # Parse "WPe2e4(x+*)" etc. _MOVE_RE = re.compile( r"^(?P[WB])" r"(?P[PNBRQK])" r"(?P[a-h][1-8])" r"(?P[a-h][1-8])" r"(?P.*)$" ) # Promotions like "=Q" or "=q" _PROMO_RE = re.compile(r"=([QRBNqrbn])") def _parse_suffix(suffix: str) -> Tuple[bool, bool, bool, Optional[str], Optional[str]]: """ Returns: is_capture, is_check, is_mate, castle_kind, promo_piece castle_kind: "k" (kingside) or "q" (queenside) or None promo_piece: one of "q","r","b","n" or None """ if not suffix: return False, False, False, None, None # Normalize suf = suffix.strip() is_capture = "x" in suf is_check = "+" in suf # Mate indicator # We'll treat any "*" as mate. is_mate = "*" in suf # Castling: dataset uses (o)/(O) in the move string for king moves castle_kind = None if "(O)" in suf: castle_kind = "q" elif "(o)" in suf: castle_kind = "k" promo_piece = None m = _PROMO_RE.search(suf) if m: promo_piece = m.group(1).lower() return is_capture, is_check, is_mate, castle_kind, promo_piece def _reindex_vocab(vocab: Dict[str, int]) -> Dict[str, int]: # sort by old id for stability items = sorted(vocab.items(), key=lambda kv: kv[1]) return {tok: new_id for new_id, (tok, _) in enumerate(items)} class ChessTokenizer(PreTrainedTokenizer): """ A custom tokenizer for chess moves using atomic decomposition. This tokenizer maps each move component to a unique token ID. The vocabulary is fixed and small (84 tokens), saving parameters. Example: >>> tokenizer = ChessTokenizer() >>> tokenizer.encode("WPe2e4 BPe7e5") [1, 4, 5, 44, 46, 4, 5, 47, 45, 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 SIDE_TOKENS = ("[W]", "[B]") PIECE_TOKENS = ("[P]", "[N]", "[B]", "[R]", "[Q]", "[K]") # flags FLAG_TOKENS = ( "[x]", # capture "[+]", # check "[#]", # mate "[O-O]", # kingside castle marker (not required by evaluator) "[O-O-O]", # queenside castle marker # promotions "[=q]", "[=r]", "[=b]", "[=n]", ) 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 the atomic vocabulary (always the same, 84 tokens) self._vocab = self._create_default_vocab() self._vocab = _reindex_vocab(self._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 the atomic vocabulary with component tokens. Total: 4 (special) + 2 (sides) + 6 (pieces) + 64 (squares) + 8 (flags) = 84 tokens """ tokens: List[str] = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN] tokens += list(self.SIDE_TOKENS) tokens += list(self.PIECE_TOKENS) # Squares (64) for file in "abcdefgh": for rank in "12345678": tokens.append(f"[{file}{rank}]") tokens += list(self.FLAG_TOKENS) return {tok: idx for idx, tok in enumerate(tokens)} @classmethod def build_vocab_from_iterator( cls, iterator, min_frequency: int = 1, ) -> "ChessTokenizer": """Build vocab (not needed for atomic approach, vocab is fixed).""" 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 = 1, max_samples: Optional[int] = None, ) -> "ChessTokenizer": """Build vocab (not needed for atomic approach, vocab is fixed).""" 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]: """ Tokenize a string of moves into a list of atomic tokens. Args: text: A string of space-separated moves. Returns: List of atomic move tokens. """ text = (text or "").strip() if not text: return [] chunks = text.split() out: List[str] = [] for chunk in chunks: # If chunk is pure uci like "e2e4" or "e7e8q" if re.fullmatch(r"[a-h][1-8][a-h][1-8][qrbn]?", chunk): src = chunk[0:2] dst = chunk[2:4] out.append(f"[{src}]") out.append(f"[{dst}]") if len(chunk) == 5 and chunk[4] in "qrbn": out.append(f"[={chunk[4]}]") continue m = _MOVE_RE.match(chunk) if not m: out.append(self.UNK_TOKEN) continue side = "[W]" if m.group("side") == "W" else "[B]" piece = m.group("piece") src = m.group("src") dst = m.group("dst") suffix = m.group("suffix") or "" out.append(side) out.append(f"[{piece}]") out.append(f"[{src}]") out.append(f"[{dst}]") is_cap, is_chk, is_mate, castle_kind, promo = _parse_suffix(suffix) # Castling markers (optional; evaluator doesn't need them) if castle_kind == "k": out.append("[O-O]") elif castle_kind == "q": out.append("[O-O-O]") if is_cap: out.append("[x]") if is_mate: out.append("[#]") elif is_chk: out.append("[+]") if promo in ("q", "r", "b", "n"): out.append(f"[={promo}]") 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,)