Chess Challenge submission by raphael-mathiot

config.json

@@ -2,10 +2,6 @@
   "architectures": [
     "ChessForCausalLM"
   ],
-  "auto_map": {
-    "AutoConfig": "model.ChessConfig",
-    "AutoModelForCausalLM": "model.ChessForCausalLM"
-  },
   "bos_token_id": 1,
   "dropout": 0.1,
   "dtype": "float32",

tokenizer.py

@@ -17,7 +17,6 @@ import json
 import os
 from pathlib import Path
 from typing import Dict, List, Optional
-import re
 
 from transformers import PreTrainedTokenizer
 
@@ -26,8 +25,14 @@ class ChessTokenizer(PreTrainedTokenizer):
     """
     A custom tokenizer for chess moves using extended UCI notation.
     
-    This tokenizer splits moves into semantic components (Pieces, Squares, Metadata).
-    Example: "WPe2e4" -> ["WP", "e2", "e4"]
+    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"]
@@ -47,6 +52,11 @@ class ChessTokenizer(PreTrainedTokenizer):
     ):
         """
         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
@@ -54,24 +64,13 @@ class ChessTokenizer(PreTrainedTokenizer):
         self._eos_token = self.EOS_TOKEN
         self._unk_token = self.UNK_TOKEN
 
-        # Clean kwargs
+        # 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)
         
-        # Regex for splitting moves into:
-        # 1. Castling: (O), (o)
-        # 2. Metadata: (x), (+*), (+)
-        # 3. Pieces: WP, BR, etc.
-        # 4. Squares: a1, h8, etc.
-        self.token_pattern = re.compile(
-            r'\(O\)|\(o\)|'        # Castling
-            r'\(x\)|\(\+\*\)|\(\+\)|'  # Metadata (Capture, Mate, Check)
-            r'[WB][PRNBQK]|'       # Pieces (Color + Type)
-            r'[a-h][1-8]'          # Squares
-        )
-
         # Load or create vocabulary
         if vocab is not None:
             self._vocab = vocab
@@ -79,13 +78,14 @@ class ChessTokenizer(PreTrainedTokenizer):
             with open(vocab_file, "r", encoding="utf-8") as f:
                 self._vocab = json.load(f)
         else:
-            # In this version, the default vocab is the FULL vocab 
-            # because chess rules are static.
+            # 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,
@@ -96,61 +96,53 @@ class ChessTokenizer(PreTrainedTokenizer):
     
     def _create_default_vocab(self) -> Dict[str, int]:
         """
-        Create the full static vocabulary for Chess.
-        Since the 'rules' of the tokens are known (squares a1-h8, pieces),
-        we generate the full map here instead of learning it.
+        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.
         """
-        # 1. Special Tokens
         special_tokens = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]
         vocab = {token: idx for idx, token in enumerate(special_tokens)}
-        idx = len(vocab)
-
-        # 2. Pieces (White/Black + Pawn/Rook/Knight/Bishop/Queen/King)
-        colors = ['W', 'B']
-        pieces = ['P', 'R', 'N', 'B', 'Q', 'K']
-        for c in colors:
-            for p in pieces:
-                token = f"{c}{p}"
-                if token not in vocab:
-                    vocab[token] = idx
-                    idx += 1
-
-        # 3. Squares (a1 to h8)
-        files = 'abcdefgh'
-        ranks = '12345678'
-        for f in files:
-            for r in ranks:
-                token = f"{f}{r}"
-                if token not in vocab:
-                    vocab[token] = idx
-                    idx += 1
-
-        # 4. Special Move Suffixes
-        # Note: Order is handled by regex, but we just need them in vocab here
-        specials = ['(O)', '(o)', '(x)', '(+)', '(+*)']
-        for s in specials:
-            if s not in vocab:
-                vocab[s] = idx
-                idx += 1
-                
         return vocab
     
     @classmethod
     def build_vocab_from_iterator(
         cls,
-        iterator: Iterator,
+        iterator,
         min_frequency: int = 1,
     ) -> "ChessTokenizer":
         """
-        API Compatibility Method.
+        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.
         
-        Since this tokenizer uses a static vocabulary based on Chess rules,
-        scanning the iterator is not necessary. We simply consume the iterator
-        (optional) and return the standard tokenizer.
+        Returns:
+            A ChessTokenizer with the built vocabulary.
         """
-        # We explicitly ignore the iterator data because our vocab 
-        # is pre-defined by the rules of the game.
-        return cls()
+        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(
@@ -162,12 +154,30 @@ class ChessTokenizer(PreTrainedTokenizer):
         max_samples: Optional[int] = 100000,
     ) -> "ChessTokenizer":
         """
-        API Compatibility Method.
+        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 tokenizer with the standard chess vocabulary.
-        Does not download the dataset as the vocabulary is static.
+        Returns:
+            A ChessTokenizer with the built vocabulary.
         """
-        return cls()
+        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:
@@ -180,71 +190,29 @@ class ChessTokenizer(PreTrainedTokenizer):
     
     def _tokenize(self, text: str) -> List[str]:
         """
-        Tokenize a string of moves into semantic components using Regex.
+        Tokenize a string of moves into a list of tokens.
         
         Args:
-            text: A string of space-separated moves (e.g., "WPe2e4 BPe7e5")
+            text: A string of space-separated moves.
         
         Returns:
-            List of components (e.g., ["WP", "e2", "e4", "BP", "e7", "e5"])
+            List of move tokens.
         """
-        # findall will ignore spaces and return only the matching components
-        return self.token_pattern.findall(text)
+        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))
+        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 _is_start_of_move(self, token: str) -> bool:
-            """
-            Helper to determine if a token represents the start of a new move.
-            Moves start with a Piece (e.g., 'WP') or Castling (e.g., '(O)').
-            """
-            # 1. Check for Castling (Short or Long)
-            if token in ['(O)', '(o)']:
-                return True
-                
-            # 2. Check for Pieces (Length 2, starts with W/B, ends with Piece type)
-            # We check specific characters to avoid confusion with squares or suffixes
-            if len(token) == 2 and token[0] in 'WB' and token[1] in 'PRNBQK':
-                return True
-                
-            return False
-
     def convert_tokens_to_string(self, tokens: List[str]) -> str:
-        """
-        Converts a list of tokens back to a string, respecting Chess notation rules.
-        
-        Logic:
-        - Spaces are inserted BEFORE a token ONLY if that token marks the start of a new move.
-        - Squares (e2, e4) and Suffixes (x, +) are concatenated to the previous token.
-        """
-        output = []
-        special_tokens = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}
-        
-        for i, token in enumerate(tokens):
-            # 1. Handle Special Tokens (keep them, surround with spaces if needed)
-            if token in special_tokens:
-                if output and output[-1] != " ":
-                    output.append(" ")
-                output.append(token)
-                
-            # 2. Handle Start of New Move (Insert space before)
-            elif self._is_start_of_move(token):
-                # Add a space if we aren't at the very start and the previous char isn't already a space
-                if output and output[-1] != " ":
-                    output.append(" ")
-                output.append(token)
-                
-            # 3. Handle Continuations (Squares 'e2', Suffixes '(x)') -> Concatenate
-            else:
-                output.append(token)
-        
-        return "".join(output).strip()
+        """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,

tokenizer_config.json

@@ -33,6 +33,12 @@
       "special": true
     }
   },
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenizer.ChessTokenizer",
+      null
+    ]
+  },
   "bos_token": "[BOS]",
   "clean_up_tokenization_spaces": false,
   "eos_token": "[EOS]",