Chess Challenge submission by Kevin-Wardakhan

README.md

@@ -0,0 +1,26 @@
+---
+library_name: transformers
+tags:
+- chess
+- llm-course
+- chess-challenge
+license: mit
+---
+
+# chess-leg-v2
+
+Chess model submitted to the LLM Course Chess Challenge.
+
+## Submission Info
+
+- **Submitted by**: [Kevin-Wardakhan](https://huggingface.co/Kevin-Wardakhan)
+- **Parameters**: 965,952
+- **Organization**: LLM-course
+
+## Model Details
+
+- **Architecture**: Chess Transformer (GPT-style)
+- **Vocab size**: 84
+- **Embedding dim**: 96
+- **Layers**: 10
+- **Heads**: 8

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": 96,
+  "n_head": 8,
+  "n_inner": 288,
+  "n_layer": 10,
+  "pad_token_id": 0,
+  "tie_weights": true,
+  "transformers_version": "4.57.6",
+  "vocab_size": 84
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39a81efaed9ce8e9ffc45c6cb1712a59305505223224b22e349643fe84f510e7
+size 3874216

special_tokens_map.json

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

tokenizer.py

@@ -0,0 +1,421 @@
+"""
+Custom Chess Tokenizer for the Chess Challenge.
+
+This tokenizer uses a STRUCTURED approach to tokenize chess moves, breaking down
+each move into its components to help the model learn legal chess patterns.
+
+The dataset format uses extended UCI notation:
+- 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
+
+Instead of treating each move as a single token (which creates thousands of tokens),
+we tokenize the COMPONENTS:
+- Color tokens: W, B
+- Piece tokens: P, N, B, R, Q, K  
+- Square tokens: a1, a2, ..., h8 (64 squares)
+- Suffix tokens: (x), (+), (+*), (o), (O), =Q, =R, =B, =N
+
+This gives ~80 tokens total, helping the model learn:
+1. Valid squares on the board
+2. Which pieces can make which types of moves
+3. The structure of legal chess moves
+"""
+
+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
+
+
+class ChessTokenizer(PreTrainedTokenizer):
+    """
+    A structured tokenizer for chess moves using component-based tokenization.
+    
+    Instead of treating each move as a single token, this tokenizer breaks moves
+    into their structural components (color, piece, from-square, to-square, suffix).
+    This smaller vocabulary helps the model learn valid chess patterns.
+    
+    Vocabulary (~80 tokens):
+    - Special: [PAD], [BOS], [EOS], [UNK]
+    - Colors: W, B
+    - Pieces: P, N, B, R, Q, K
+    - Squares: a1-h8 (64 tokens)
+    - Suffixes: (x), (+), (+*), (o), (O), =Q, =R, =B, =N
+    
+    Example:
+        >>> tokenizer = ChessTokenizer()
+        >>> tokens = tokenizer.tokenize("WPe2e4 BPe7e5")
+        >>> print(tokens)
+        ['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]"
+    
+    # Chess components
+    COLORS = ["W", "B"]
+    PIECES = ["P", "N", "B", "R", "Q", "K"]
+    FILES = ["a", "b", "c", "d", "e", "f", "g", "h"]
+    RANKS = ["1", "2", "3", "4", "5", "6", "7", "8"]
+    SQUARES = [f + r for f in ["a", "b", "c", "d", "e", "f", "g", "h"] 
+               for r in ["1", "2", "3", "4", "5", "6", "7", "8"]]  # a1, a2, ..., h8
+    SUFFIXES = ["(x)", "(+)", "(+*)", "(o)", "(O)", "=Q", "=R", "=B", "=N"]
+    
+    # Regex pattern to parse extended UCI moves
+    # Format: [W|B][Piece][from_sq][to_sq][optional: =PromoPiece][optional: suffix]
+    MOVE_PATTERN = re.compile(
+        r'^([WB])([PNBRQK])([a-h][1-8])([a-h][1-8])(=[QRBN])?(\([xo+*O]+\))?$'
+    )
+    
+    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 structured vocabulary
+            self._vocab = self._create_structured_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_structured_vocab(self) -> Dict[str, int]:
+        """
+        Create the structured vocabulary with all chess components.
+        
+        This creates a fixed vocabulary of ~85 tokens covering all possible
+        chess move components.
+        """
+        tokens = []
+        
+        # Special tokens first
+        tokens.extend([self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN])
+        
+        # Colors
+        tokens.extend(self.COLORS)
+        
+        # Pieces
+        tokens.extend(self.PIECES)
+        
+        # Squares (64 tokens)
+        tokens.extend(self.SQUARES)
+        
+        # Suffixes
+        tokens.extend(self.SUFFIXES)
+        
+        # Build vocabulary
+        vocab = {token: idx for idx, token in enumerate(tokens)}
+        return vocab
+    
+    def _create_default_vocab(self) -> Dict[str, int]:
+        """Alias for _create_structured_vocab for compatibility."""
+        return self._create_structured_vocab()
+    
+    def _parse_move(self, move: str) -> List[str]:
+        """
+        Parse a single move into its component tokens.
+        
+        Args:
+            move: A move in extended UCI format (e.g., "WPe2e4", "BNg8f6(x)").
+        
+        Returns:
+            List of component tokens.
+        """
+        move = move.strip()
+        if not move:
+            return []
+        
+        # Handle special tokens
+        if move in [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]:
+            return [move]
+        
+        # Try to match the move pattern
+        match = self.MOVE_PATTERN.match(move)
+        if match:
+            color, piece, from_sq, to_sq, promotion, suffix = match.groups()
+            tokens = [color, piece, from_sq, to_sq]
+            if promotion:
+                tokens.append(promotion)
+            if suffix:
+                tokens.append(suffix)
+            return tokens
+        
+        # If pattern doesn't match, try to extract what we can
+        # This handles edge cases and malformed moves gracefully
+        tokens = []
+        i = 0
+        
+        # Color (W or B)
+        if i < len(move) and move[i] in self.COLORS:
+            tokens.append(move[i])
+            i += 1
+        
+        # Piece (P, N, B, R, Q, K)
+        if i < len(move) and move[i] in self.PIECES:
+            tokens.append(move[i])
+            i += 1
+        
+        # From square (e.g., e2)
+        if i + 1 < len(move) and move[i:i+2] in self.SQUARES:
+            tokens.append(move[i:i+2])
+            i += 2
+        
+        # To square (e.g., e4)
+        if i + 1 < len(move) and move[i:i+2] in self.SQUARES:
+            tokens.append(move[i:i+2])
+            i += 2
+        
+        # Promotion (e.g., =Q)
+        if i + 1 < len(move) and move[i:i+2] in self.SUFFIXES:
+            tokens.append(move[i:i+2])
+            i += 2
+        
+        # Suffix (e.g., (x), (+), (+*), (o), (O))
+        remaining = move[i:]
+        if remaining in self.SUFFIXES:
+            tokens.append(remaining)
+        elif remaining:
+            # Try to find a matching suffix
+            for suffix in self.SUFFIXES:
+                if remaining.startswith(suffix):
+                    tokens.append(suffix)
+                    break
+        
+        # If we couldn't parse anything, return UNK
+        if not tokens:
+            return [self.UNK_TOKEN]
+        
+        return tokens
+    
+    @classmethod
+    def build_vocab_from_iterator(
+        cls,
+        iterator,
+        min_frequency: int = 1,
+    ) -> "ChessTokenizer":
+        """
+        Build a tokenizer (for compatibility - vocab is fixed).
+        
+        The structured tokenizer has a fixed vocabulary, so this method
+        simply returns a new tokenizer instance.
+        
+        Args:
+            iterator: An iterator yielding game strings (ignored for structured vocab).
+            min_frequency: Minimum frequency (ignored for structured vocab).
+        
+        Returns:
+            A ChessTokenizer with the structured 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 (for compatibility - vocab is fixed).
+        
+        The structured tokenizer has a fixed vocabulary covering all valid
+        chess move components, so no dataset scanning is needed.
+        
+        Args:
+            dataset_name: Name of the dataset (ignored).
+            split: Dataset split (ignored).
+            column: Column name (ignored).
+            min_frequency: Minimum frequency (ignored).
+            max_samples: Maximum samples (ignored).
+        
+        Returns:
+            A ChessTokenizer with the structured 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]:
+        """
+        Tokenize a string of moves into component tokens.
+        
+        Args:
+            text: A string of space-separated moves.
+        
+        Returns:
+            List of component tokens.
+        """
+        tokens = []
+        moves = text.strip().split()
+        
+        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, 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 move string.
+        
+        Reconstructs moves from component tokens by grouping them appropriately.
+        """
+        # Filter out special tokens
+        special = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}
+        tokens = [t for t in tokens if t not in special]
+        
+        if not tokens:
+            return ""
+        
+        # Reconstruct moves from components
+        result = []
+        current_move = []
+        
+        for token in tokens:
+            # Start of a new move (color token)
+            if token in self.COLORS:
+                if current_move:
+                    result.append("".join(current_move))
+                current_move = [token]
+            else:
+                current_move.append(token)
+        
+        # Don't forget the last move
+        if current_move:
+            result.append("".join(current_move))
+        
+        return " ".join(result)
+    
+    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).
+    
+    With the structured tokenizer, this counts component frequencies.
+    
+    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
+    
+    tokenizer = ChessTokenizer()
+    
+    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:
+        tokens = tokenizer._tokenize(example[column])
+        token_counts.update(tokens)
+    
+    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
+    },
+    "3": {
+      "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 +1,86 @@
+{
+  "[PAD]": 0,
+  "[BOS]": 1,
+  "[EOS]": 2,
+  "[UNK]": 3,
+  "W": 4,
+  "B": 8,
+  "P": 6,
+  "N": 7,
+  "R": 9,
+  "Q": 10,
+  "K": 11,
+  "a1": 12,
+  "a2": 13,
+  "a3": 14,
+  "a4": 15,
+  "a5": 16,
+  "a6": 17,
+  "a7": 18,
+  "a8": 19,
+  "b1": 20,
+  "b2": 21,
+  "b3": 22,
+  "b4": 23,
+  "b5": 24,
+  "b6": 25,
+  "b7": 26,
+  "b8": 27,
+  "c1": 28,
+  "c2": 29,
+  "c3": 30,
+  "c4": 31,
+  "c5": 32,
+  "c6": 33,
+  "c7": 34,
+  "c8": 35,
+  "d1": 36,
+  "d2": 37,
+  "d3": 38,
+  "d4": 39,
+  "d5": 40,
+  "d6": 41,
+  "d7": 42,
+  "d8": 43,
+  "e1": 44,
+  "e2": 45,
+  "e3": 46,
+  "e4": 47,
+  "e5": 48,
+  "e6": 49,
+  "e7": 50,
+  "e8": 51,
+  "f1": 52,
+  "f2": 53,
+  "f3": 54,
+  "f4": 55,
+  "f5": 56,
+  "f6": 57,
+  "f7": 58,
+  "f8": 59,
+  "g1": 60,
+  "g2": 61,
+  "g3": 62,
+  "g4": 63,
+  "g5": 64,
+  "g6": 65,
+  "g7": 66,
+  "g8": 67,
+  "h1": 68,
+  "h2": 69,
+  "h3": 70,
+  "h4": 71,
+  "h5": 72,
+  "h6": 73,
+  "h7": 74,
+  "h8": 75,
+  "(x)": 76,
+  "(+)": 77,
+  "(+*)": 78,
+  "(o)": 79,
+  "(O)": 80,
+  "=Q": 81,
+  "=R": 82,
+  "=B": 83,
+  "=N": 84
+}