Chess Challenge submission by agroudiev

README.md

@@ -14,13 +14,13 @@ Chess model submitted to the LLM Course Chess Challenge.
 ## Submission Info
 
 - **Submitted by**: [agroudiev](https://huggingface.co/agroudiev)
-- **Parameters**: 999,896
+- **Parameters**: 990,004
 - **Organization**: LLM-course
 
 ## Model Details
 
 - **Architecture**: Chess Transformer (GPT-style)
-- **Vocab size**: 139
+- **Vocab size**: 134
 - **Embedding dim**: 128
 - **Layers**: 6
-- **Heads**: 8
+- **Heads**: 4

config.json

@@ -10,11 +10,11 @@
   "model_type": "chess_transformer",
   "n_ctx": 256,
   "n_embd": 128,
-  "n_head": 8,
-  "n_inner": 356,
+  "n_head": 4,
+  "n_inner": 350,
   "n_layer": 6,
   "pad_token_id": 0,
   "tie_weights": true,
   "transformers_version": "4.57.4",
-  "vocab_size": 139
+  "vocab_size": 134
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b7c76abb2502ca91c2fb9acfb939e55e6015ef8e30bd347133665ab0838cf33e
-size 4006032
+oid sha256:24c1decf60ef24cda76b5d1e328823253e4d38bdc08dfae6ad569aed71570aa5
+size 3966464

tokenizer.py

@@ -0,0 +1,278 @@
+"""
+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,)
+
+
+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

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

vocab.json

@@ -58,84 +58,79 @@
   "BPg6g5": 56,
   "BPg7g5": 57,
   "BPg7g6": 58,
-  "BPh5h4": 59,
-  "BPh7h5": 60,
-  "BPh7h6": 61,
-  "BQd8b6": 62,
-  "BQd8c7": 63,
-  "BQd8d7": 64,
-  "BQd8e7": 65,
-  "BRa8b8": 66,
-  "BRa8c8": 67,
-  "BRa8d8": 68,
-  "BRf8d8": 69,
-  "BRf8e8": 70,
-  "WBc1b2": 71,
-  "WBc1d2": 72,
-  "WBc1e3": 73,
-  "WBc1f4": 74,
-  "WBc1g5": 75,
-  "WBf1b5": 76,
-  "WBf1c4": 77,
-  "WBf1d3": 78,
-  "WBf1e2": 79,
-  "WBf1g2": 80,
-  "WBg5f6(x)": 81,
-  "WKe1c1(O)": 82,
-  "WKe1g1(o)": 83,
-  "WKg1g2": 84,
-  "WKg1h1": 85,
-  "WKg1h2": 86,
-  "WNb1c3": 87,
-  "WNb1d2": 88,
-  "WNc3d5": 89,
-  "WNf3d4(x)": 90,
-  "WNf3e5": 91,
-  "WNf3e5(x)": 92,
-  "WNf3g5": 93,
-  "WNg1e2": 94,
-  "WNg1f3": 95,
-  "WPa2a3": 96,
-  "WPa2a4": 97,
-  "WPa4a5": 98,
-  "WPb2b3": 99,
-  "WPb2b4": 100,
-  "WPb2c3(x)": 101,
-  "WPb4b5": 102,
-  "WPc2c3": 103,
-  "WPc2c4": 104,
-  "WPc3c4": 105,
-  "WPc3d4(x)": 106,
-  "WPc4c5": 107,
-  "WPc4d5(x)": 108,
-  "WPd2d3": 109,
-  "WPd2d4": 110,
-  "WPd3d4": 111,
-  "WPd4d5": 112,
-  "WPd4e5(x)": 113,
-  "WPe2e3": 114,
-  "WPe2e4": 115,
-  "WPe3e4": 116,
-  "WPe4d5(x)": 117,
-  "WPe4e5": 118,
-  "WPf2f3": 119,
-  "WPf2f4": 120,
-  "WPf4f5": 121,
-  "WPg2g3": 122,
-  "WPg2g4": 123,
-  "WPg4g5": 124,
-  "WPh2h3": 125,
-  "WPh2h4": 126,
-  "WPh3h4": 127,
-  "WPh4h5": 128,
-  "WQd1c2": 129,
-  "WQd1d2": 130,
-  "WQd1e2": 131,
-  "WQd1f3": 132,
-  "WRa1b1": 133,
-  "WRa1c1": 134,
-  "WRa1d1": 135,
-  "WRa1e1": 136,
-  "WRf1d1": 137,
-  "WRf1e1": 138
+  "BPh7h5": 59,
+  "BPh7h6": 60,
+  "BQd8b6": 61,
+  "BQd8c7": 62,
+  "BQd8d7": 63,
+  "BQd8e7": 64,
+  "BRa8b8": 65,
+  "BRa8c8": 66,
+  "BRa8d8": 67,
+  "BRf8e8": 68,
+  "WBc1b2": 69,
+  "WBc1d2": 70,
+  "WBc1e3": 71,
+  "WBc1f4": 72,
+  "WBc1g5": 73,
+  "WBf1b5": 74,
+  "WBf1c4": 75,
+  "WBf1d3": 76,
+  "WBf1e2": 77,
+  "WBf1g2": 78,
+  "WKe1c1(O)": 79,
+  "WKe1g1(o)": 80,
+  "WKg1g2": 81,
+  "WKg1h1": 82,
+  "WKg1h2": 83,
+  "WNb1c3": 84,
+  "WNb1d2": 85,
+  "WNc3d5": 86,
+  "WNf3d4(x)": 87,
+  "WNf3e5": 88,
+  "WNf3e5(x)": 89,
+  "WNf3g5": 90,
+  "WNg1e2": 91,
+  "WNg1f3": 92,
+  "WPa2a3": 93,
+  "WPa2a4": 94,
+  "WPa4a5": 95,
+  "WPb2b3": 96,
+  "WPb2b4": 97,
+  "WPb2c3(x)": 98,
+  "WPb4b5": 99,
+  "WPc2c3": 100,
+  "WPc2c4": 101,
+  "WPc3c4": 102,
+  "WPc3d4(x)": 103,
+  "WPc4c5": 104,
+  "WPc4d5(x)": 105,
+  "WPd2d3": 106,
+  "WPd2d4": 107,
+  "WPd3d4": 108,
+  "WPd4d5": 109,
+  "WPd4e5(x)": 110,
+  "WPe2e3": 111,
+  "WPe2e4": 112,
+  "WPe3e4": 113,
+  "WPe4d5(x)": 114,
+  "WPe4e5": 115,
+  "WPf2f3": 116,
+  "WPf2f4": 117,
+  "WPf4f5": 118,
+  "WPg2g3": 119,
+  "WPg2g4": 120,
+  "WPg4g5": 121,
+  "WPh2h3": 122,
+  "WPh2h4": 123,
+  "WPh3h4": 124,
+  "WPh4h5": 125,
+  "WQd1c2": 126,
+  "WQd1d2": 127,
+  "WQd1e2": 128,
+  "WRa1b1": 129,
+  "WRa1c1": 130,
+  "WRa1d1": 131,
+  "WRa1e1": 132,
+  "WRf1e1": 133
 }