Chess Challenge submission by etiennemaugars

README.md

@@ -0,0 +1,26 @@
+---
+library_name: transformers
+tags:
+- chess
+- llm-course
+- chess-challenge
+license: mit
+---
+
+# emaugars_chess_dmd
+
+Chess model submitted to the LLM Course Chess Challenge.
+
+## Submission Info
+
+- **Submitted by**: [etiennemaugars](https://huggingface.co/etiennemaugars)
+- **Parameters**: 953,728
+- **Organization**: LLM-course
+
+## Model Details
+
+- **Architecture**: Chess Transformer (GPT-style)
+- **Vocab size**: 148
+- **Embedding dim**: 128
+- **Layers**: 6
+- **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": 320,
+  "n_embd": 128,
+  "n_head": 4,
+  "n_inner": 320,
+  "n_layer": 6,
+  "pad_token_id": 0,
+  "tie_weights": true,
+  "transformers_version": "4.57.6",
+  "vocab_size": 148
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f898a29640d37703dfc7c4555d8e8c54001efdde85c4847fc07e2259a2314d4
+size 3821360

special_tokens_map.json

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

tokenizer.py

@@ -0,0 +1,446 @@
+"""
+Custom Chess Tokenizer for the Chess Challenge.
+
+This tokenizer treats each move as a single token using the extended UCI notation
+from the Lichess dataset (e.g., WPe2e4, BNg8f6).
+
+The dataset format uses:
+- 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
+"""
+
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+from typing import Dict, List, Optional
+
+from transformers import PreTrainedTokenizer
+
+
+class ChessTokenizer(PreTrainedTokenizer):
+    """
+    A custom tokenizer for chess moves using extended UCI notation.
+
+    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"]
+    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 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.
+        """
+        special_tokens = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]
+        vocab = {token: idx for idx, token in enumerate(special_tokens)}
+        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.
+
+        Returns:
+            A ChessTokenizer with the built vocabulary.
+        """
+        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(
+        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 (default: 500).
+            max_samples: Maximum number of samples to process (default: 100k).
+
+        Returns:
+            A ChessTokenizer with the built vocabulary.
+        """
+        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:
+        """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 tokens.
+
+        Args:
+            text: A string of space-separated moves.
+
+        Returns:
+            List of move tokens.
+        """
+        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, 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,)
+
+
+class DecomposedChessTokenizer(PreTrainedTokenizer):
+    """
+    Tokenizer that decomposes moves into: [Piece] [FromSquare] [ToSquare] [Promotion?]
+    """
+
+    # Special tokens
+    PAD_TOKEN = "[PAD]"
+    BOS_TOKEN = "[BOS]"
+    EOS_TOKEN = "[EOS]"
+    UNK_TOKEN = "[UNK]"
+
+    def __init__(self, **kwargs):
+        self._pad_token = self.PAD_TOKEN
+        self._bos_token = self.BOS_TOKEN
+        self._eos_token = self.EOS_TOKEN
+        self._unk_token = self.UNK_TOKEN
+
+        kwargs.pop("pad_token", None)
+        kwargs.pop("bos_token", None)
+        kwargs.pop("eos_token", None)
+        kwargs.pop("unk_token", None)
+
+        # Build vocabulary
+        self._vocab = self._build_decomposed_vocab()
+        self._ids_to_tokens = {v: k for k, v in self._vocab.items()}
+
+        super().__init__(
+            pad_token=self._pad_token,
+            bos_token=self._bos_token,
+            eos_token=self._eos_token,
+            unk_token=self._unk_token,
+            **kwargs,
+        )
+
+    def _build_decomposed_vocab(self) -> Dict[str, int]:
+        vocab = {}
+        idx = 0
+
+        # Special tokens
+        for token in [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]:
+            vocab[token] = idx
+            idx += 1
+
+        # Piece tokens: WP, WN, WB, WR, WQ, WK, BP, BN, BB, BR, BQ, BK
+        for color in ["W", "B"]:
+            for piece in ["P", "N", "B", "R", "Q", "K"]:
+                vocab[f"{color}{piece}"] = idx
+                idx += 1
+
+        # Square tokens with suffixes
+        files = "abcdefgh"
+        ranks = "12345678"
+        for file in files:
+            for rank in ranks:
+                # From square suffix
+                vocab[f"{file}{rank}_f"] = idx
+                idx += 1
+                # To square suffix
+                vocab[f"{file}{rank}_t"] = idx
+                idx += 1
+
+        # Promotion tokens
+        for piece in ["Q", "R", "B", "N"]:
+            vocab[f"={piece}"] = idx
+            idx += 1
+
+        return vocab
+
+    def _tokenize(self, text: str) -> List[str]:
+        """
+        Tokenize moves in format: "WPe2e4 BNg8f6"
+        Output: ["WP", "e2_f", "e4_t", "BN", "g8_f", "f6_t"]
+        """
+        tokens = []
+        moves = text.strip().split()
+
+        for move in moves:
+            if move in [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]:
+                tokens.append(move)
+                continue
+
+            # Parse extended UCI format: [W|B][Piece][from][to][annotations]
+            if len(move) < 6:
+                tokens.append(self.UNK_TOKEN)
+                continue
+
+            color_piece = move[:2]  # e.g., "WP"
+            from_sq = move[2:4]  # e.g., "e2"
+            to_sq = move[4:6]  # e.g., "e4"
+
+            tokens.append(color_piece)
+            tokens.append(f"{from_sq}_f")
+            tokens.append(f"{to_sq}_t")
+
+            # Check for promotion
+            if "=" in move:
+                promo_idx = move.index("=")
+                if promo_idx + 1 < len(move):
+                    promo_piece = move[promo_idx + 1]
+                    if promo_piece in "QRBN":
+                        tokens.append(f"={promo_piece}")
+
+        return tokens
+
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        """Convert decomposed tokens back to move string for display."""
+        result = []
+        i = 0
+        while i < len(tokens):
+            token = tokens[i]
+
+            # Skip special tokens
+            if token in [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]:
+                i += 1
+                continue
+
+            # Check if this starts a move (piece token)
+            if len(token) == 2 and token[0] in "WB" and token[1] in "PNBRQK":
+                move_parts = [token]
+                # Expect from_sq, to_sq, optional promotion
+                for j in range(1, 4):
+                    if i + j < len(tokens):
+                        next_token = tokens[i + j]
+                        if next_token.endswith("_f") or next_token.endswith("_t"):
+                            move_parts.append(next_token.replace("_f", "").replace("_t", ""))
+                        elif next_token.startswith("="):
+                            move_parts.append(next_token)
+                        else:
+                            break
+                    else:
+                        break
+
+                # Format: WP + e2 + e4 -> WPe2e4
+                if len(move_parts) >= 3:
+                    result.append("".join(move_parts))
+                    i += len(move_parts)
+                else:
+                    i += 1
+            else:
+                i += 1
+
+        return " ".join(result)
+
+    @property
+    def vocab_size(self) -> int:
+        return len(self._vocab)
+
+    def get_vocab(self) -> Dict[str, int]:
+        return dict(self._vocab)
+
+    def _convert_token_to_id(self, token: str) -> int:
+        return self._vocab.get(token, self._vocab.get(self.UNK_TOKEN, 0))
+
+    def _convert_id_to_token(self, index: int) -> str:
+        return self._ids_to_tokens.get(index, self.UNK_TOKEN)
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> tuple:
+        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
+    },
+    "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 +1,150 @@
+{
+  "[PAD]": 0,
+  "[BOS]": 1,
+  "[EOS]": 2,
+  "[UNK]": 3,
+  "WP": 4,
+  "WN": 5,
+  "WB": 6,
+  "WR": 7,
+  "WQ": 8,
+  "WK": 9,
+  "BP": 10,
+  "BN": 11,
+  "BB": 12,
+  "BR": 13,
+  "BQ": 14,
+  "BK": 15,
+  "a1_f": 16,
+  "a1_t": 17,
+  "a2_f": 18,
+  "a2_t": 19,
+  "a3_f": 20,
+  "a3_t": 21,
+  "a4_f": 22,
+  "a4_t": 23,
+  "a5_f": 24,
+  "a5_t": 25,
+  "a6_f": 26,
+  "a6_t": 27,
+  "a7_f": 28,
+  "a7_t": 29,
+  "a8_f": 30,
+  "a8_t": 31,
+  "b1_f": 32,
+  "b1_t": 33,
+  "b2_f": 34,
+  "b2_t": 35,
+  "b3_f": 36,
+  "b3_t": 37,
+  "b4_f": 38,
+  "b4_t": 39,
+  "b5_f": 40,
+  "b5_t": 41,
+  "b6_f": 42,
+  "b6_t": 43,
+  "b7_f": 44,
+  "b7_t": 45,
+  "b8_f": 46,
+  "b8_t": 47,
+  "c1_f": 48,
+  "c1_t": 49,
+  "c2_f": 50,
+  "c2_t": 51,
+  "c3_f": 52,
+  "c3_t": 53,
+  "c4_f": 54,
+  "c4_t": 55,
+  "c5_f": 56,
+  "c5_t": 57,
+  "c6_f": 58,
+  "c6_t": 59,
+  "c7_f": 60,
+  "c7_t": 61,
+  "c8_f": 62,
+  "c8_t": 63,
+  "d1_f": 64,
+  "d1_t": 65,
+  "d2_f": 66,
+  "d2_t": 67,
+  "d3_f": 68,
+  "d3_t": 69,
+  "d4_f": 70,
+  "d4_t": 71,
+  "d5_f": 72,
+  "d5_t": 73,
+  "d6_f": 74,
+  "d6_t": 75,
+  "d7_f": 76,
+  "d7_t": 77,
+  "d8_f": 78,
+  "d8_t": 79,
+  "e1_f": 80,
+  "e1_t": 81,
+  "e2_f": 82,
+  "e2_t": 83,
+  "e3_f": 84,
+  "e3_t": 85,
+  "e4_f": 86,
+  "e4_t": 87,
+  "e5_f": 88,
+  "e5_t": 89,
+  "e6_f": 90,
+  "e6_t": 91,
+  "e7_f": 92,
+  "e7_t": 93,
+  "e8_f": 94,
+  "e8_t": 95,
+  "f1_f": 96,
+  "f1_t": 97,
+  "f2_f": 98,
+  "f2_t": 99,
+  "f3_f": 100,
+  "f3_t": 101,
+  "f4_f": 102,
+  "f4_t": 103,
+  "f5_f": 104,
+  "f5_t": 105,
+  "f6_f": 106,
+  "f6_t": 107,
+  "f7_f": 108,
+  "f7_t": 109,
+  "f8_f": 110,
+  "f8_t": 111,
+  "g1_f": 112,
+  "g1_t": 113,
+  "g2_f": 114,
+  "g2_t": 115,
+  "g3_f": 116,
+  "g3_t": 117,
+  "g4_f": 118,
+  "g4_t": 119,
+  "g5_f": 120,
+  "g5_t": 121,
+  "g6_f": 122,
+  "g6_t": 123,
+  "g7_f": 124,
+  "g7_t": 125,
+  "g8_f": 126,
+  "g8_t": 127,
+  "h1_f": 128,
+  "h1_t": 129,
+  "h2_f": 130,
+  "h2_t": 131,
+  "h3_f": 132,
+  "h3_t": 133,
+  "h4_f": 134,
+  "h4_t": 135,
+  "h5_f": 136,
+  "h5_t": 137,
+  "h6_f": 138,
+  "h6_t": 139,
+  "h7_f": 140,
+  "h7_t": 141,
+  "h8_f": 142,
+  "h8_t": 143,
+  "=Q": 144,
+  "=R": 145,
+  "=B": 146,
+  "=N": 147
+}