Chess Challenge submission by thandre10

README.md

@@ -0,0 +1,26 @@
+---
+library_name: transformers
+tags:
+- chess
+- llm-course
+- chess-challenge
+license: mit
+---
+
+# chess_thandre10_v3
+
+Chess model submitted to the LLM Course Chess Challenge.
+
+## Submission Info
+
+- **Submitted by**: [thandre10](https://huggingface.co/thandre10)
+- **Parameters**: 738,360
+- **Organization**: LLM-course
+
+## Model Details
+
+- **Architecture**: Chess Transformer (GPT-style)
+- **Vocab size**: 83
+- **Embedding dim**: 120
+- **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": 120,
+  "n_head": 4,
+  "n_inner": 480,
+  "n_layer": 4,
+  "pad_token_id": 0,
+  "tie_weights": true,
+  "transformers_version": "4.57.6",
+  "vocab_size": 83
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8de0af4b0f13d2a81c8fa50f28bd6718b3d44f685324ece2ef2e70cfa80bd521
+size 2957840

special_tokens_map.json

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

tokenizer.py

@@ -0,0 +1,412 @@
+"""
+Custom Chess Tokenizer for the Chess Challenge.
+
+This tokenizer uses sub-structural tokenization: each move is decomposed into
+its components (piece, source square, destination square, suffix) instead of
+treating the whole move as a single token.
+
+Example: WPe2e4 -> [P, e2, e4]  (color is implicit from move number)
+         BNg8f6(x) -> [N, g8, f6, (x)]
+
+This approach:
+- Reduces vocabulary from ~1200 to ~80 tokens
+- Enables generalization across similar moves
+- Eliminates [UNK] tokens for rare moves
+- Saves parameters in the embedding layer
+
+The dataset format uses:
+- W/B prefix for White/Black (ignored - implicit from position)
+- 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
+"""
+
+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
+
+
+# Regex pattern to parse extended UCI notation
+# Matches: (W|B)(Piece)(src_file)(src_rank)(dst_file)(dst_rank)(suffix?)
+MOVE_PATTERN = re.compile(
+    r'^([WB])([PNBRQK])([a-h])([1-8])([a-h])([1-8])(\([^)]+\))?$'
+)
+
+
+class ChessTokenizer(PreTrainedTokenizer):
+    """
+    A custom tokenizer for chess moves using sub-structural tokenization.
+
+    Each move is decomposed into components:
+    - Piece type (P, N, B, R, Q, K)
+    - Source square (e2, d7, etc.)
+    - Destination square (e4, f6, etc.)
+    - Optional suffix for captures/checks ((x), (+), (+*), (o), (O))
+
+    The color (W/B) is NOT tokenized as it's implicit from the move order.
+
+    Example:
+        >>> tokenizer = ChessTokenizer.build_vocab()
+        >>> tokenizer.encode("WPe2e4 BPe7e5")
+        [1, 5, 20, 28, 5, 52, 44, 2]  # [BOS, P, e2, e4, 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]"
+    
+    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 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,
+            eos_token=self._eos_token,
+            unk_token=self._unk_token,
+            **kwargs,
+        )
+    
+    def _create_default_vocab(self) -> Dict[str, int]:
+        """
+        Create the full sub-structural vocabulary.
+
+        The vocabulary contains:
+        - 4 special tokens: [PAD], [BOS], [EOS], [UNK]
+        - 6 piece tokens: P, N, B, R, Q, K
+        - 64 square tokens: a1, a2, ..., h8
+        - 5 suffix tokens: (x), (+), (+*), (o), (O)
+
+        Total: 79 tokens (vs ~1200 for move-level tokenization)
+        """
+        tokens = []
+
+        # Special tokens first
+        special_tokens = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]
+        tokens.extend(special_tokens)
+
+        # Piece tokens
+        pieces = ['P', 'N', 'B', 'R', 'Q', 'K']
+        tokens.extend(pieces)
+
+        # Square tokens (a1-h8)
+        files = 'abcdefgh'
+        ranks = '12345678'
+        for f in files:
+            for r in ranks:
+                tokens.append(f + r)
+
+        # Suffix tokens for special moves
+        suffixes = ['(x)', '(+)', '(+*)', '(o)', '(O)']
+        tokens.extend(suffixes)
+
+        # Promotion tokens (pawn promotion to piece)
+        # Format in dataset might be like WPe7e8Q for promotion
+        promotion_pieces = ['=Q', '=R', '=B', '=N']
+        tokens.extend(promotion_pieces)
+
+        vocab = {token: idx for idx, token in enumerate(tokens)}
+        return vocab
+    
+    @classmethod
+    def build_vocab(cls) -> "ChessTokenizer":
+        """
+        Build a tokenizer with the pre-defined sub-structural vocabulary.
+
+        This is the recommended way to create a tokenizer for the chess challenge.
+        The vocabulary is deterministic and covers all possible moves.
+
+        Returns:
+            A ChessTokenizer with the full sub-structural vocabulary (~83 tokens).
+        """
+        return cls()
+
+    @classmethod
+    def build_vocab_from_iterator(
+        cls,
+        iterator,
+        min_frequency: int = 1,
+    ) -> "ChessTokenizer":
+        """
+        Build a tokenizer vocabulary from an iterator of game strings.
+
+        Note: With sub-structural tokenization, this method is mainly useful
+        for analyzing token frequencies. The default vocabulary already covers
+        all possible moves.
+
+        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.
+        """
+        # With sub-structural tokenization, we use the default vocab
+        # which already contains all possible sub-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.
+
+        Note: With sub-structural tokenization, the vocabulary is pre-defined
+        and doesn't need to be built from data. This method is kept for
+        compatibility but simply returns a tokenizer with the default vocab.
+
+        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.
+            max_samples: Maximum number of samples to process.
+
+        Returns:
+            A ChessTokenizer with the full sub-structural vocabulary.
+        """
+        # With sub-structural tokenization, we don't need to scan the dataset
+        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 _parse_move(self, move: str) -> List[str]:
+        """
+        Parse a single move into its sub-components.
+
+        Args:
+            move: A move in extended UCI notation (e.g., WPe2e4, BNg8f6(x))
+
+        Returns:
+            List of tokens: [piece, src_square, dst_square, suffix?]
+            Color (W/B) is ignored as it's implicit from move order.
+        """
+        # Try standard move pattern
+        match = MOVE_PATTERN.match(move)
+        if match:
+            color, piece, src_file, src_rank, dst_file, dst_rank, suffix = match.groups()
+            tokens = [piece, src_file + src_rank, dst_file + dst_rank]
+            if suffix:
+                tokens.append(suffix)
+            return tokens
+
+        # Try promotion pattern: WPe7e8Q or WPe7e8Q(+)
+        promo_pattern = re.match(
+            r'^([WB])P([a-h])([1-8])([a-h])([1-8])([QRBN])(\([^)]+\))?$',
+            move
+        )
+        if promo_pattern:
+            color, src_file, src_rank, dst_file, dst_rank, promo_piece, suffix = promo_pattern.groups()
+            tokens = ['P', src_file + src_rank, dst_file + dst_rank, '=' + promo_piece]
+            if suffix:
+                tokens.append(suffix)
+            return tokens
+
+        # Fallback: return as single token (will likely be UNK)
+        return [move]
+
+    def _tokenize(self, text: str) -> List[str]:
+        """
+        Tokenize a string of moves into sub-structural tokens.
+
+        Each move is decomposed into:
+        - Piece type (P, N, B, R, Q, K)
+        - Source square (e2, d7, etc.)
+        - Destination square (e4, f6, etc.)
+        - Optional suffix ((x), (+), etc.)
+
+        Args:
+            text: A string of space-separated moves.
+
+        Returns:
+            List of sub-tokens.
+
+        Example:
+            "WPe2e4 BPe7e5" -> ['P', 'e2', 'e4', 'P', 'e7', 'e5']
+        """
+        tokens = []
+        moves = text.strip().split()
+        for move in moves:
+            tokens.extend(self._parse_move(move))
+        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 sub-tokens back to a string of moves.
+
+        Reconstructs moves from their components. Each move consists of:
+        - Piece token (P, N, B, R, Q, K)
+        - Source square (e2, d7, etc.)
+        - Destination square (e4, f6, etc.)
+        - Optional suffix ((x), (+), etc.) or promotion (=Q, =R, etc.)
+
+        Args:
+            tokens: List of sub-tokens.
+
+        Returns:
+            Space-separated string of reconstructed moves.
+        """
+        special = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}
+        pieces = {'P', 'N', 'B', 'R', 'Q', 'K'}
+        suffixes = {'(x)', '(+)', '(+*)', '(o)', '(O)'}
+        promotions = {'=Q', '=R', '=B', '=N'}
+
+        moves = []
+        current_move = []
+
+        for token in tokens:
+            if token in special:
+                continue
+
+            if token in pieces:
+                # Start of a new move - save previous if exists
+                if current_move:
+                    moves.append(''.join(current_move))
+                current_move = [token]
+            elif token in suffixes or token in promotions:
+                # End of move with suffix/promotion
+                current_move.append(token)
+            else:
+                # Square token
+                current_move.append(token)
+
+        # Don't forget the last move
+        if current_move:
+            moves.append(''.join(current_move))
+
+        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 sub-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 sub-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))))
+
+    # Use a tokenizer instance to parse moves into sub-tokens
+    tokenizer = ChessTokenizer()
+    token_counts = Counter()
+
+    for example in dataset:
+        sub_tokens = tokenizer._tokenize(example[column])
+        token_counts.update(sub_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 +1,85 @@
+{
+  "[PAD]": 0,
+  "[BOS]": 1,
+  "[EOS]": 2,
+  "[UNK]": 3,
+  "P": 4,
+  "N": 5,
+  "B": 6,
+  "R": 7,
+  "Q": 8,
+  "K": 9,
+  "a1": 10,
+  "a2": 11,
+  "a3": 12,
+  "a4": 13,
+  "a5": 14,
+  "a6": 15,
+  "a7": 16,
+  "a8": 17,
+  "b1": 18,
+  "b2": 19,
+  "b3": 20,
+  "b4": 21,
+  "b5": 22,
+  "b6": 23,
+  "b7": 24,
+  "b8": 25,
+  "c1": 26,
+  "c2": 27,
+  "c3": 28,
+  "c4": 29,
+  "c5": 30,
+  "c6": 31,
+  "c7": 32,
+  "c8": 33,
+  "d1": 34,
+  "d2": 35,
+  "d3": 36,
+  "d4": 37,
+  "d5": 38,
+  "d6": 39,
+  "d7": 40,
+  "d8": 41,
+  "e1": 42,
+  "e2": 43,
+  "e3": 44,
+  "e4": 45,
+  "e5": 46,
+  "e6": 47,
+  "e7": 48,
+  "e8": 49,
+  "f1": 50,
+  "f2": 51,
+  "f3": 52,
+  "f4": 53,
+  "f5": 54,
+  "f6": 55,
+  "f7": 56,
+  "f8": 57,
+  "g1": 58,
+  "g2": 59,
+  "g3": 60,
+  "g4": 61,
+  "g5": 62,
+  "g6": 63,
+  "g7": 64,
+  "g8": 65,
+  "h1": 66,
+  "h2": 67,
+  "h3": 68,
+  "h4": 69,
+  "h5": 70,
+  "h6": 71,
+  "h7": 72,
+  "h8": 73,
+  "(x)": 74,
+  "(+)": 75,
+  "(+*)": 76,
+  "(o)": 77,
+  "(O)": 78,
+  "=Q": 79,
+  "=R": 80,
+  "=B": 81,
+  "=N": 82
+}