Add chess model submission

README.md

@@ -0,0 +1,55 @@
+---
+library_name: transformers
+tags:
+- chess
+- llm-course
+- chess-challenge
+license: mit
+---
+
+# chess-ooooooooo
+
+A chess transformer model trained for the LLM Course Chess Challenge.
+
+## Model Architecture
+
+This model uses a GPT-style transformer architecture optimized for chess move prediction:
+
+- **Parameters**: 948,352 (0.95M)
+- **Vocabulary size**: 85
+- **Embedding dimension**: 128
+- **Number of layers**: 6
+- **Attention heads**: 4
+- **Feed-forward dimension**: 320
+- **Context length**: 256
+- **Dropout**: 0.1
+
+## Training
+
+The model was trained on a subset of the Lichess 2025 dataset, focusing on learning valid chess move sequences. The architecture was carefully tuned to stay within the 1M parameter constraint while maintaining reasonable performance.
+
+## Usage
+
+```python
+from transformers import AutoModelForCausalLM
+from src.tokenizer import ChessTokenizer
+
+model = AutoModelForCausalLM.from_pretrained(
+    "LLM-course/chess-ooooooooo",
+    trust_remote_code=True
+)
+tokenizer = ChessTokenizer.from_pretrained(
+    "LLM-course/chess-ooooooooo",
+    trust_remote_code=True
+)
+
+# Generate moves
+input_text = "[BOS] WPe2e4"
+input_ids = tokenizer.encode(input_text)
+outputs = model.generate(input_ids, max_length=50)
+predicted_moves = tokenizer.decode(outputs[0])
+```
+
+## Submission
+
+Submitted by [etienneLefranc](https://huggingface.co/etienneLefranc) for the LLM Course Chess Challenge.

config.json

@@ -0,0 +1,21 @@
+{
+  "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": 128,
+  "n_head": 4,
+  "n_inner": 320,
+  "n_layer": 6,
+  "pad_token_id": 0,
+  "tie_weights": false,
+  "tie_word_embeddings": false,
+  "transformers_version": "4.57.6",
+  "vocab_size": 85
+}

model.py

@@ -0,0 +1,436 @@
+"""
+Chess Transformer Model for the Chess Challenge.
+
+This module provides a simple GPT-style transformer architecture
+designed to fit within the 1M parameter constraint.
+
+Key components:
+- ChessConfig: Configuration class for model hyperparameters
+- ChessForCausalLM: The main model class for next-move prediction
+"""
+
+from __future__ import annotations
+
+import math
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import PretrainedConfig, PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+
+class ChessConfig(PretrainedConfig):
+    """
+    Configuration class for the Chess Transformer model.
+    
+    This configuration is designed for a ~1M parameter model.
+    Students can adjust these values to explore different architectures.
+    
+    Parameter budget breakdown (with default values):
+    - Embeddings (vocab): 1200 x 128 = 153,600
+    - Position Embeddings: 256 x 128 = 32,768
+    - Transformer Layers: 6 x ~120,000 = ~720,000
+    - LM Head (with weight tying): 0 (shared with embeddings)
+    - Total: ~906,000 parameters
+    
+    Attributes:
+        vocab_size: Size of the vocabulary (number of unique moves).
+        n_embd: Embedding dimension (d_model).
+        n_layer: Number of transformer layers.
+        n_head: Number of attention heads.
+        n_ctx: Maximum sequence length (context window).
+        n_inner: Feed-forward inner dimension (default: 3 * n_embd).
+        dropout: Dropout probability.
+        layer_norm_epsilon: Epsilon for layer normalization.
+        tie_weights: Whether to tie embedding and output weights.
+    """
+    
+    model_type = "chess_transformer"
+    
+    def __init__(
+        self,
+        vocab_size: int = 1200,
+        n_embd: int = 128,
+        n_layer: int = 6,
+        n_head: int = 4,
+        n_ctx: int = 256,
+        n_inner: Optional[int] = None,
+        dropout: float = 0.1,
+        layer_norm_epsilon: float = 1e-5,
+        tie_weights: bool = True,
+        pad_token_id: int = 0,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        **kwargs,
+    ):
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            **kwargs,
+        )
+        
+        self.vocab_size = vocab_size
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.n_ctx = n_ctx
+        self.n_inner = n_inner if n_inner is not None else 3 * n_embd  # Reduced from 4x to 3x
+        self.dropout = dropout
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.tie_weights = tie_weights
+        # Inform HF base class about tying behavior
+        self.tie_word_embeddings = bool(tie_weights)
+
+
+class MultiHeadAttention(nn.Module):
+    """
+    Multi-head self-attention module.
+    
+    This is a standard scaled dot-product attention implementation
+    with causal masking for autoregressive generation.
+    """
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        
+        assert config.n_embd % config.n_head == 0, \
+            f"n_embd ({config.n_embd}) must be divisible by n_head ({config.n_head})"
+        
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.head_dim = config.n_embd // config.n_head
+        
+        # Combined QKV projection for efficiency
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd)
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd)
+        
+        self.dropout = nn.Dropout(config.dropout)
+        
+        # Causal mask (will be created on first forward pass)
+        self.register_buffer(
+            "bias",
+            torch.tril(torch.ones(config.n_ctx, config.n_ctx)).view(
+                1, 1, config.n_ctx, config.n_ctx
+            ),
+            persistent=False,
+        )
+    
+    def forward(
+        self,
+        x: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, seq_len, _ = x.size()
+        
+        # Compute Q, K, V
+        qkv = self.c_attn(x)
+        q, k, v = qkv.split(self.n_embd, dim=2)
+        
+        # Reshape for multi-head attention
+        q = q.view(batch_size, seq_len, self.n_head, self.head_dim).transpose(1, 2)
+        k = k.view(batch_size, seq_len, self.n_head, self.head_dim).transpose(1, 2)
+        v = v.view(batch_size, seq_len, self.n_head, self.head_dim).transpose(1, 2)
+        
+        # Scaled dot-product attention
+        attn_weights = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
+        
+        # Apply causal mask
+        causal_mask = self.bias[:, :, :seq_len, :seq_len]
+        attn_weights = attn_weights.masked_fill(causal_mask == 0, float("-inf"))
+        
+        # Apply attention mask (for padding)
+        if attention_mask is not None:
+            # attention_mask shape: (batch_size, seq_len) -> (batch_size, 1, 1, seq_len)
+            attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+            attn_weights = attn_weights.masked_fill(attention_mask == 0, float("-inf"))
+        
+        attn_weights = F.softmax(attn_weights, dim=-1)
+        attn_weights = self.dropout(attn_weights)
+        
+        # Apply attention to values
+        attn_output = torch.matmul(attn_weights, v)
+        
+        # Reshape back
+        attn_output = attn_output.transpose(1, 2).contiguous().view(
+            batch_size, seq_len, self.n_embd
+        )
+        
+        # Output projection
+        attn_output = self.c_proj(attn_output)
+        
+        return attn_output
+
+
+class FeedForward(nn.Module):
+    """
+    Feed-forward network (MLP) module.
+    
+    Standard two-layer MLP with GELU activation.
+    """
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        
+        self.c_fc = nn.Linear(config.n_embd, config.n_inner)
+        self.c_proj = nn.Linear(config.n_inner, config.n_embd)
+        self.dropout = nn.Dropout(config.dropout)
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.c_fc(x)
+        x = F.gelu(x)
+        x = self.c_proj(x)
+        x = self.dropout(x)
+        return x
+
+
+class TransformerBlock(nn.Module):
+    """
+    A single transformer block with attention and feed-forward layers.
+    
+    Uses pre-normalization (LayerNorm before attention/FFN) for better
+    training stability.
+    """
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        
+        self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.attn = MultiHeadAttention(config)
+        self.ln_2 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.mlp = FeedForward(config)
+    
+    def forward(
+        self,
+        x: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        # Pre-norm attention
+        x = x + self.attn(self.ln_1(x), attention_mask=attention_mask)
+        # Pre-norm FFN
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class ChessForCausalLM(PreTrainedModel):
+    """
+    Chess Transformer for Causal Language Modeling (next-move prediction).
+    
+    This model is designed to predict the next chess move given a sequence
+    of previous moves. It uses a GPT-style architecture with:
+    - Token embeddings for chess moves
+    - Learned positional embeddings
+    - Stacked transformer blocks
+    - Linear head for next-token prediction
+    
+    The model supports weight tying between the embedding layer and the
+    output projection to save parameters.
+    
+    Example:
+        >>> config = ChessConfig(vocab_size=1200, n_embd=128, n_layer=6)
+        >>> model = ChessForCausalLM(config)
+        >>> inputs = {"input_ids": torch.tensor([[1, 42, 87]])}
+        >>> outputs = model(**inputs)
+        >>> next_move_logits = outputs.logits[:, -1, :]
+    """
+    
+    config_class = ChessConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    # Suppress missing-key warning for tied lm_head when loading
+    keys_to_ignore_on_load_missing = ["lm_head.weight"]
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__(config)
+        
+        # Token and position embeddings
+        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
+        self.wpe = nn.Embedding(config.n_ctx, config.n_embd)
+        
+        self.drop = nn.Dropout(config.dropout)
+        
+        # Transformer blocks
+        self.h = nn.ModuleList([
+            TransformerBlock(config) for _ in range(config.n_layer)
+        ])
+        
+        # Final layer norm
+        self.ln_f = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        
+        # Output head
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+        
+        # Initialize weights first
+        self.post_init()
+        
+        # Tie weights if configured (after post_init)
+        # transformers will handle _tied_weights_keys automatically via tie_weights()
+        if config.tie_weights:
+            self.tie_weights()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.wte
+
+    def set_input_embeddings(self, new_embeddings: nn.Module):
+        self.wte = new_embeddings
+        if getattr(self.config, "tie_weights", False):
+            self.tie_weights()
+
+    def get_output_embeddings(self) -> nn.Module:
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings: nn.Module):
+        self.lm_head = new_embeddings
+
+    def tie_weights(self, recompute_mapping=False, **kwargs):
+        # Tie weights between embedding and output layers
+        if getattr(self.config, "tie_weights", False) or getattr(self.config, "tie_word_embeddings", False):
+            # Directly tie the weights
+            self.lm_head.weight = self.wte.weight
+    
+    def _init_weights(self, module: nn.Module):
+        """Initialize weights following GPT-2 style."""
+        if isinstance(module, nn.Linear):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
+            if module.bias is not None:
+                torch.nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
+        elif isinstance(module, nn.LayerNorm):
+            torch.nn.init.ones_(module.weight)
+            torch.nn.init.zeros_(module.bias)
+    
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        """
+        Forward pass of the model.
+        
+        Args:
+            input_ids: Token IDs of shape (batch_size, seq_len).
+            attention_mask: Attention mask of shape (batch_size, seq_len).
+            position_ids: Position IDs of shape (batch_size, seq_len).
+            labels: Labels for language modeling loss.
+            return_dict: Whether to return a ModelOutput object.
+        
+        Returns:
+            CausalLMOutputWithPast containing loss (if labels provided) and logits.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        
+        batch_size, seq_len = input_ids.size()
+        device = input_ids.device
+        
+        # Create position IDs if not provided
+        if position_ids is None:
+            position_ids = torch.arange(seq_len, device=device).unsqueeze(0).expand(batch_size, -1)
+        
+        # Get embeddings
+        token_embeds = self.wte(input_ids)
+        position_embeds = self.wpe(position_ids)
+        hidden_states = self.drop(token_embeds + position_embeds)
+        
+        # Pass through transformer blocks
+        for block in self.h:
+            hidden_states = block(hidden_states, attention_mask=attention_mask)
+        
+        # Final layer norm
+        hidden_states = self.ln_f(hidden_states)
+        
+        # Get logits
+        logits = self.lm_head(hidden_states)
+        
+        # Compute loss if labels are provided
+        loss = None
+        if labels is not None:
+            # Shift logits and labels for next-token prediction
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            
+            # Flatten for cross-entropy
+            loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
+            # loss_fct = nn.CrossEntropyLoss(ignore_index=self.config.pad_token_id)
+            loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)),
+                shift_labels.view(-1),
+            )
+        
+        if not return_dict:
+            output = (logits,)
+            return ((loss,) + output) if loss is not None else output
+        
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=None,
+            hidden_states=None,
+            attentions=None,
+        )
+    
+    @torch.no_grad()
+    def generate_move(
+        self,
+        input_ids: torch.LongTensor,
+        temperature: float = 1.0,
+        top_k: Optional[int] = None,
+        top_p: Optional[float] = None,
+    ) -> int:
+        """
+        Generate the next move given a sequence of moves.
+        
+        Args:
+            input_ids: Token IDs of shape (1, seq_len).
+            temperature: Sampling temperature (1.0 = no change).
+            top_k: If set, only sample from top k tokens.
+            top_p: If set, use nucleus sampling with this threshold.
+        
+        Returns:
+            The token ID of the predicted next move.
+        """
+        self.eval()
+        
+        # Get logits for the last position
+        outputs = self(input_ids)
+        logits = outputs.logits[:, -1, :] / temperature
+        
+        # Apply top-k filtering
+        if top_k is not None:
+            indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+            logits[indices_to_remove] = float("-inf")
+        
+        # Apply top-p (nucleus) filtering
+        if top_p is not None:
+            sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+            cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+            
+            # Remove tokens with cumulative probability above the threshold
+            sorted_indices_to_remove = cumulative_probs > top_p
+            sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+            sorted_indices_to_remove[..., 0] = 0
+            
+            indices_to_remove = sorted_indices_to_remove.scatter(
+                dim=-1, index=sorted_indices, src=sorted_indices_to_remove
+            )
+            logits[indices_to_remove] = float("-inf")
+        
+        # Sample from the distribution
+        probs = F.softmax(logits, dim=-1)
+        next_token = torch.multinomial(probs, num_samples=1)
+        
+        return next_token.item()
+
+
+# Register the model with Auto classes for easy loading
+from transformers import AutoConfig, AutoModelForCausalLM
+
+AutoConfig.register("chess_transformer", ChessConfig)
+AutoModelForCausalLM.register(ChessConfig, ChessForCausalLM)

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eec45323a73f8c60461b30c496cb3326e229e97dfb328f562bfc121581b24068
+size 3799936

special_tokens_map.json

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

tokenizer.py

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

@@ -0,0 +1,53 @@
+{
+  "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
+    }
+  },
+  "additional_special_tokens": null,
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenizer.ChessTokenizer",
+      null
+    ]
+  },
+  "backend": "custom",
+  "bos_token": "[BOS]",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "[EOS]",
+  "is_local": true,
+  "model_max_length": 1000000000000000019884624838656,
+  "model_specific_special_tokens": {},
+  "pad_token": "[PAD]",
+  "tokenizer_class": "ChessTokenizer",
+  "unk_token": "[UNK]"
+}

vocab.json

@@ -0,0 +1,87 @@
+{
+  "[PAD]": 0,
+  "[BOS]": 1,
+  "[EOS]": 2,
+  "[UNK]": 3,
+  "[W]": 4,
+  "[B]": 5,
+  "[P]": 6,
+  "[N]": 7,
+  "[BISHOP]": 8,
+  "[R]": 9,
+  "[Q]": 10,
+  "[K]": 11,
+  "[a1]": 12,
+  "[b1]": 13,
+  "[c1]": 14,
+  "[d1]": 15,
+  "[e1]": 16,
+  "[f1]": 17,
+  "[g1]": 18,
+  "[h1]": 19,
+  "[a2]": 20,
+  "[b2]": 21,
+  "[c2]": 22,
+  "[d2]": 23,
+  "[e2]": 24,
+  "[f2]": 25,
+  "[g2]": 26,
+  "[h2]": 27,
+  "[a3]": 28,
+  "[b3]": 29,
+  "[c3]": 30,
+  "[d3]": 31,
+  "[e3]": 32,
+  "[f3]": 33,
+  "[g3]": 34,
+  "[h3]": 35,
+  "[a4]": 36,
+  "[b4]": 37,
+  "[c4]": 38,
+  "[d4]": 39,
+  "[e4]": 40,
+  "[f4]": 41,
+  "[g4]": 42,
+  "[h4]": 43,
+  "[a5]": 44,
+  "[b5]": 45,
+  "[c5]": 46,
+  "[d5]": 47,
+  "[e5]": 48,
+  "[f5]": 49,
+  "[g5]": 50,
+  "[h5]": 51,
+  "[a6]": 52,
+  "[b6]": 53,
+  "[c6]": 54,
+  "[d6]": 55,
+  "[e6]": 56,
+  "[f6]": 57,
+  "[g6]": 58,
+  "[h6]": 59,
+  "[a7]": 60,
+  "[b7]": 61,
+  "[c7]": 62,
+  "[d7]": 63,
+  "[e7]": 64,
+  "[f7]": 65,
+  "[g7]": 66,
+  "[h7]": 67,
+  "[a8]": 68,
+  "[b8]": 69,
+  "[c8]": 70,
+  "[d8]": 71,
+  "[e8]": 72,
+  "[f8]": 73,
+  "[g8]": 74,
+  "[h8]": 75,
+  "[x]": 76,
+  "[+]": 77,
+  "[#]": 78,
+  "[O-O]": 79,
+  "[O-O-O]": 80,
+  "[prom_Q]": 81,
+  "[prom_R]": 82,
+  "[prom_B]": 83,
+  "[prom_N]": 84
+}