Chess Challenge submission by Valbad

README.md

@@ -0,0 +1,26 @@
+---
+library_name: transformers
+tags:
+- chess
+- llm-course
+- chess-challenge
+license: mit
+---
+
+# chess-baseline-valbad
+
+Chess model submitted to the LLM Course Chess Challenge.
+
+## Submission Info
+
+- **Submitted by**: [Valbad](https://huggingface.co/Valbad)
+- **Parameters**: 930,600
+- **Organization**: LLM-course
+
+## Model Details
+
+- **Architecture**: Chess Transformer (GPT-style)
+- **Vocab size**: 5000
+- **Embedding dim**: 100
+- **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": 100,
+  "n_head": 4,
+  "n_inner": 300,
+  "n_layer": 4,
+  "pad_token_id": 0,
+  "tie_weights": true,
+  "transformers_version": "4.57.3",
+  "vocab_size": 5000
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7ff1be405662ed91f0c3feba04fa9ca40ce87f870891a8a787e838f9a25e420
+size 3726776

special_tokens_map.json

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

tokenizer.py

@@ -0,0 +1,370 @@
+"""
+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_iterator(
+        cls,
+        iterator,
+        vocab_size: int = 1200,
+        min_frequency: int = 1,
+        ) -> "ChessTokenizer":
+        """
+        Build a tokenizer vocabulary from an iterator of game strings.
+
+        - Controls final vocab size explicitly via vocab_size.
+        - Keeps the most frequent move tokens (best coverage).
+        - Uses min_frequency as a floor, but vocab_size is the main control.
+        """
+        from collections import Counter
+
+        token_counts = Counter()
+        for game in iterator:
+            moves = game.strip().split()
+            token_counts.update(moves)
+
+        # Filter by min_frequency first
+        items = [(tok, cnt) for tok, cnt in token_counts.items() if cnt >= min_frequency]
+
+        # Sort by frequency desc, then token for determinism
+        items.sort(key=lambda x: (-x[1], x[0]))
+
+        special_tokens = [cls.PAD_TOKEN, cls.BOS_TOKEN, cls.EOS_TOKEN, cls.UNK_TOKEN]
+        max_move_tokens = max(0, vocab_size - len(special_tokens))
+
+        move_tokens = [tok for tok, _ in items[:max_move_tokens]]
+        vocab = {token: idx for idx, token in enumerate(special_tokens + move_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)
+    @classmethod
+    def build_vocab_from_dataset(
+        cls,
+        dataset_name: str = "dlouapre/lichess_2025-01_1M",
+        split: str = "train",
+        column: str = "text",
+        vocab_size: int = 1200,
+        min_frequency: int = 1,
+        max_samples: Optional[int] = 200000,
+    ) -> "ChessTokenizer":
+        """
+        Build a tokenizer vocabulary from a Hugging Face dataset.
+
+        Args:
+            vocab_size: Final vocab size INCLUDING special tokens.
+            min_frequency: Minimum count to consider a move (usually 1 is fine).
+            max_samples: How many games to scan to build vocab.
+        """
+        from datasets import load_dataset
+
+        dataset = load_dataset(dataset_name, split=split)
+
+        # if max_samples is not None: # v0&1
+        #     dataset = dataset.select(range(min(max_samples, len(dataset))))
+         
+        if max_samples is not None: # v2
+            n = min(max_samples, len(dataset))
+            dataset = dataset.shuffle(seed=42).select(range(n))
+
+        def game_iterator():
+            for example in dataset:
+                yield example[column]
+
+        return cls.build_vocab_from_iterator(
+            game_iterator(),
+            vocab_size=vocab_size,
+            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 build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+    #     if token_ids_1 is not None:
+    #         # Not expected here, but handle gracefully
+    #         token_ids = token_ids_0 + token_ids_1
+    #     else:
+    #         token_ids = token_ids_0
+    #     return [self.bos_token_id] + token_ids + [self.eos_token_id]
+
+    # def get_special_tokens_mask(self, token_ids_0, token_ids_1=None, already_has_special_tokens=False):
+    #     if already_has_special_tokens:
+    #         return [1 if t in (self.pad_token_id, self.bos_token_id, self.eos_token_id, self.unk_token_id) else 0 for t in token_ids_0]
+    #     if token_ids_1 is not None:
+    #         token_ids = token_ids_0 + token_ids_1
+    #     else:
+    #         token_ids = token_ids_0
+    #     return [1] + [0] * len(token_ids) + [1]
+
+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]"
+}