Chess Challenge submission by gagliardidf

README.md

@@ -0,0 +1,26 @@
+---
+library_name: transformers
+tags:
+- chess
+- llm-course
+- chess-challenge
+license: mit
+---
+
+# gagliardidf-chess-model-v2
+
+Chess model submitted to the LLM Course Chess Challenge.
+
+## Submission Info
+
+- **Submitted by**: [gagliardidf](https://huggingface.co/gagliardidf)
+- **Parameters**: 836,992
+- **Organization**: LLM-course
+
+## Model Details
+
+- **Architecture**: Chess Transformer (GPT-style)
+- **Vocab size**: 85
+- **Embedding dim**: 128
+- **Layers**: 4
+- **Heads**: 4

config.json

@@ -0,0 +1,20 @@
+{
+  "architectures": [
+    "ChessForCausalLM"
+  ],
+  "bos_token_id": 1,
+  "dropout": 0.1,
+  "dtype": "float32",
+  "eos_token_id": 2,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "chess_transformer",
+  "n_ctx": 256,
+  "n_embd": 128,
+  "n_head": 4,
+  "n_inner": 512,
+  "n_layer": 4,
+  "pad_token_id": 0,
+  "tie_weights": true,
+  "transformers_version": "4.57.6",
+  "vocab_size": 85
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f5a9a13b659d52d0a4907503f7fe38b8ac8cd58f45682cbfb465e9422a33bbbe
+size 3352376

special_tokens_map.json

@@ -0,0 +1,6 @@
+{
+  "bos_token": "[BOS]",
+  "eos_token": "[EOS]",
+  "pad_token": "[PAD]",
+  "unk_token": "[UNK]"
+}

tokenizer.py

@@ -0,0 +1,469 @@
+"""
+Custom Chess Tokenizer for the Chess Challenge.
+
+This tokenizer uses 85 tokens total:
+- 64 Squares: Each square (a1–h8) is one token
+- 6 Pieces: P, N, B, R, Q, K
+- 10 Special Indicators: Capture (x), Check (+), Checkmate (#), Kingside Castle (O-O),
+  Queenside Castle (O-O-O), and Promotion pieces (q, r, b, n)
+- 5 Control Tokens: [PAD], [BOS], [EOS], [WHITE], [BLACK]
+
+The dataset format uses extended UCI notation (e.g., WPe2e4, BNg8f6(x)):
+- W/B prefix maps to [WHITE]/[BLACK] tokens
+- Piece letter (P, N, B, R, Q, K) is tokenized
+- Source and destination squares are tokenized
+- Special suffixes are extracted and tokenized
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+from pathlib import Path
+from typing import Dict, List, Optional
+
+from transformers import PreTrainedTokenizer
+
+
+class ChessTokenizer(PreTrainedTokenizer):
+    """
+    A custom tokenizer for chess moves using 85 tokens.
+    
+    This tokenizer decomposes moves into component tokens:
+    - Color token ([WHITE] or [BLACK])
+    - Piece token (P, N, B, R, Q, K)
+    - Source square token (a1-h8)
+    - Destination square token (a1-h8)
+    - Optional special indicator tokens (x, +, #, O-O, O-O-O, q/r/b/n)
+    
+    Example:
+        >>> tokenizer = ChessTokenizer()
+        >>> tokenizer.encode("WPe2e4 BNg8f6(x)")
+        [1, 69, 65, 36, 40, 70, 66, 52, 46, 71, 2]  # [BOS, WHITE, P, e2, e4, BLACK, N, g8, f6, x, 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]"
+    WHITE_TOKEN = "[WHITE]"
+    BLACK_TOKEN = "[BLACK]"
+    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 the fixed 85-token vocabulary
+            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 the fixed 85-token vocabulary.
+        
+        Vocabulary structure:
+        - 5 Control Tokens: [PAD], [BOS], [EOS], [WHITE], [BLACK]
+        - 64 Square Tokens: a1, a2, ..., a8, b1, ..., h8
+        - 6 Piece Tokens: P, N, B, R, Q, K
+        - 10 Special Indicator Tokens: x, +, #, O-O, O-O-O, q, r, b, n, (placeholder for UNK)
+        """
+        vocab = {}
+        idx = 0
+        
+        # 5 Control Tokens
+        control_tokens = [
+            self.PAD_TOKEN,
+            self.BOS_TOKEN,
+            self.EOS_TOKEN,
+            self.WHITE_TOKEN,
+            self.BLACK_TOKEN,
+        ]
+        for token in control_tokens:
+            vocab[token] = idx
+            idx += 1
+        
+        # 64 Square Tokens (a1-h8)
+        for file in 'abcdefgh':
+            for rank in '12345678':
+                square = f"{file}{rank}"
+                vocab[square] = idx
+                idx += 1
+        
+        # 6 Piece Tokens
+        pieces = ['P', 'N', 'B', 'R', 'Q', 'K']
+        for piece in pieces:
+            vocab[piece] = idx
+            idx += 1
+        
+        # 10 Special Indicator Tokens
+        special_indicators = [
+            'x',      # Capture
+            '+',      # Check
+            '#',      # Checkmate
+            'O-O',    # Kingside castle
+            'O-O-O',  # Queenside castle
+            'q',      # Queen promotion
+            'r',      # Rook promotion
+            'b',      # Bishop promotion
+            'n',      # Knight promotion
+            self.UNK_TOKEN,  # Unknown (keeping for compatibility)
+        ]
+        for indicator in special_indicators:
+            vocab[indicator] = idx
+            idx += 1
+        
+        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 (not used, vocab is fixed).
+        
+        Returns:
+            A ChessTokenizer with the fixed 85-token vocabulary.
+        """
+        # The vocabulary is fixed: 85 tokens
+        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 (not used, vocab is fixed).
+            max_samples: Maximum number of samples to process (default: 100k).
+        
+        Returns:
+            A ChessTokenizer with the fixed 85-token vocabulary.
+        """
+        # The vocabulary is fixed: 85 tokens
+        return cls()
+    
+    @property
+    def vocab_size(self) -> int:
+        """Return the size of the vocabulary (85 tokens)."""
+        return len(self._vocab)
+    
+    def get_vocab(self) -> Dict[str, int]:
+        """Return the vocabulary as a dictionary."""
+        return dict(self._vocab)
+    
+    def _parse_move(self, move: str) -> List[str]:
+        """
+        Parse a single move string into component tokens.
+        
+        Moves are in extended UCI format: [W|B][Piece][from_sq][to_sq][suffixes]
+        Examples:
+            "WPe2e4" -> [WHITE, P, e2, e4]
+            "BNg8f6(x)" -> [BLACK, N, g8, f6, x]
+            "WPe7e8=Q(+*)" -> [WHITE, P, e7, e8, q, #]
+            "WKe1g1(o)" -> [WHITE, K, e1, g1, O-O]
+        
+        Args:
+            move: Move string in extended UCI format.
+        
+        Returns:
+            List of tokens representing the move.
+        """
+        tokens = []
+        
+        if len(move) < 6:
+            return [self.UNK_TOKEN]
+        
+        # Extract color (W or B)
+        color = move[0]
+        if color == 'W':
+            tokens.append(self.WHITE_TOKEN)
+        elif color == 'B':
+            tokens.append(self.BLACK_TOKEN)
+        else:
+            tokens.append(self.UNK_TOKEN)
+            return tokens
+        
+        # Extract piece
+        piece = move[1]
+        if piece in ['P', 'N', 'B', 'R', 'Q', 'K']:
+            tokens.append(piece)
+        else:
+            tokens.append(self.UNK_TOKEN)
+            return tokens
+        
+        # Extract source and destination squares
+        from_sq = move[2:4]
+        to_sq = move[4:6]
+        
+        # Validate squares are valid (a1-h8)
+        if from_sq[0] in 'abcdefgh' and from_sq[1] in '12345678':
+            tokens.append(from_sq)
+        else:
+            tokens.append(self.UNK_TOKEN)
+        
+        if to_sq[0] in 'abcdefgh' and to_sq[1] in '12345678':
+            tokens.append(to_sq)
+        else:
+            tokens.append(self.UNK_TOKEN)
+        
+        # Extract and parse suffixes
+        suffix_pattern = r'\(([^)]+)\)|(=.)'
+        suffixes = re.findall(suffix_pattern, move[6:])
+        
+        # Check for castling indicators (o) or (O)
+        if '(o)' in move or '(O)' in move:
+            # Determine kingside vs queenside by move
+            # Kingside: e1g1, e8g8 (short)
+            # Queenside: e1c1, e8c8 (long)
+            if to_sq in ['g1', 'g8']:
+                tokens.append('O-O')
+            elif to_sq in ['c1', 'c8']:
+                tokens.append('O-O-O')
+            else:
+                # Default to kingside if unclear
+                tokens.append('O-O')
+        else:
+            # Check for promotion (=Q, =R, =B, =N)
+            promotion_match = re.search(r'=([QRBN])', move)
+            if promotion_match:
+                promo_piece = promotion_match.group(1).lower()
+                if promo_piece in ['q', 'r', 'b', 'n']:
+                    tokens.append(promo_piece)
+            
+            # Check for capture (x)
+            if '(x)' in move or 'x' in move[6:]:
+                tokens.append('x')
+            
+            # Check for checkmate (+*)
+            if '(+*)' in move or '(x+*)' in move or '(x+*' in move:
+                tokens.append('#')
+            # Check for check (+)
+            elif '(+)' in move or '(x+)' in move:
+                tokens.append('+')
+        
+        return tokens
+    
+    def _tokenize(self, text: str) -> List[str]:
+        """
+        Tokenize a string of moves into a list of tokens.
+        
+        Each move is decomposed into its component tokens:
+        [WHITE/BLACK] + [Piece] + [from_sq] + [to_sq] + [special indicators...]
+        
+        Args:
+            text: A string of space-separated moves in extended UCI format.
+        
+        Returns:
+            List of tokens representing all moves.
+        """
+        moves = text.strip().split()
+        tokens = []
+        
+        for move in moves:
+            move_tokens = self._parse_move(move)
+            tokens.extend(move_tokens)
+        
+        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, 3))
+    
+    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.
+        
+        Reconstructs moves in simplified UCI format from component tokens.
+        
+        Args:
+            tokens: List of tokens representing moves.
+        
+        Returns:
+            String representation of moves in space-separated format.
+        """
+        # Filter out special control tokens (keep WHITE/BLACK for move boundaries)
+        control_tokens = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN}
+        filtered_tokens = [t for t in tokens if t not in control_tokens]
+        
+        moves = []
+        i = 0
+        while i < len(filtered_tokens):
+            # Look for move pattern: [WHITE/BLACK] + [Piece] + [from_sq] + [to_sq] + [indicators...]
+            move_parts = []
+            
+            # Check for color token
+            if i < len(filtered_tokens) and filtered_tokens[i] in [self.WHITE_TOKEN, self.BLACK_TOKEN]:
+                move_parts.append(filtered_tokens[i])
+                i += 1
+            
+            # Get piece
+            if i < len(filtered_tokens) and filtered_tokens[i] in ['P', 'N', 'B', 'R', 'Q', 'K']:
+                move_parts.append(filtered_tokens[i])
+                i += 1
+            else:
+                i += 1
+                continue
+            
+            # Get squares
+            if i + 1 < len(filtered_tokens):
+                from_sq = filtered_tokens[i]
+                to_sq = filtered_tokens[i + 1]
+                # Validate squares
+                if (len(from_sq) == 2 and from_sq[0] in 'abcdefgh' and from_sq[1] in '12345678' and
+                    len(to_sq) == 2 and to_sq[0] in 'abcdefgh' and to_sq[1] in '12345678'):
+                    move_parts.extend([from_sq, to_sq])
+                    i += 2
+                    
+                    # Collect special indicators
+                    while i < len(filtered_tokens):
+                        next_token = filtered_tokens[i]
+                        if next_token in ['x', '+', '#', 'O-O', 'O-O-O', 'q', 'r', 'b', 'n']:
+                            move_parts.append(next_token)
+                            i += 1
+                        else:
+                            break
+                    
+                    # Reconstruct move string (simplified UCI format)
+                    if len(move_parts) >= 4:
+                        move_str = move_parts[2] + move_parts[3]  # from_sq + to_sq
+                        # Add promotion if present
+                        for promo in ['q', 'r', 'b', 'n']:
+                            if promo in move_parts:
+                                move_str += f"={promo.upper()}"
+                                break
+                        moves.append(move_str)
+                else:
+                    i += 2
+            else:
+                break
+        
+        return " ".join(moves)
+    
+    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)

tokenizer_config.json

@@ -0,0 +1,50 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "[BOS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "[EOS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "84": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenizer.ChessTokenizer",
+      null
+    ]
+  },
+  "bos_token": "[BOS]",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "[EOS]",
+  "extra_special_tokens": {},
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "[PAD]",
+  "tokenizer_class": "ChessTokenizer",
+  "unk_token": "[UNK]"
+}

vocab.json

@@ -0,0 +1,87 @@
+{
+  "[PAD]": 0,
+  "[BOS]": 1,
+  "[EOS]": 2,
+  "[WHITE]": 3,
+  "[BLACK]": 4,
+  "a1": 5,
+  "a2": 6,
+  "a3": 7,
+  "a4": 8,
+  "a5": 9,
+  "a6": 10,
+  "a7": 11,
+  "a8": 12,
+  "b1": 13,
+  "b2": 14,
+  "b3": 15,
+  "b4": 16,
+  "b5": 17,
+  "b6": 18,
+  "b7": 19,
+  "b8": 20,
+  "c1": 21,
+  "c2": 22,
+  "c3": 23,
+  "c4": 24,
+  "c5": 25,
+  "c6": 26,
+  "c7": 27,
+  "c8": 28,
+  "d1": 29,
+  "d2": 30,
+  "d3": 31,
+  "d4": 32,
+  "d5": 33,
+  "d6": 34,
+  "d7": 35,
+  "d8": 36,
+  "e1": 37,
+  "e2": 38,
+  "e3": 39,
+  "e4": 40,
+  "e5": 41,
+  "e6": 42,
+  "e7": 43,
+  "e8": 44,
+  "f1": 45,
+  "f2": 46,
+  "f3": 47,
+  "f4": 48,
+  "f5": 49,
+  "f6": 50,
+  "f7": 51,
+  "f8": 52,
+  "g1": 53,
+  "g2": 54,
+  "g3": 55,
+  "g4": 56,
+  "g5": 57,
+  "g6": 58,
+  "g7": 59,
+  "g8": 60,
+  "h1": 61,
+  "h2": 62,
+  "h3": 63,
+  "h4": 64,
+  "h5": 65,
+  "h6": 66,
+  "h7": 67,
+  "h8": 68,
+  "P": 69,
+  "N": 70,
+  "B": 71,
+  "R": 72,
+  "Q": 73,
+  "K": 74,
+  "x": 75,
+  "+": 76,
+  "#": 77,
+  "O-O": 78,
+  "O-O-O": 79,
+  "q": 80,
+  "r": 81,
+  "b": 82,
+  "n": 83,
+  "[UNK]": 84
+}