Chess Challenge submission by Oceane28

README.md

@@ -7,20 +7,16 @@ tags:
 license: mit
 ---
 
-# chess-5-Oceane28
-
-Chess model submitted to the LLM Course Chess Challenge.
-
-## Submission Info
+## Chess model submitted to the LLM Course Chess Challenge.
 
+### Submission Info
 - **Submitted by**: [Oceane28](https://huggingface.co/Oceane28)
-- **Parameters**: 997,488
+- **Parameters**: 998,988
 - **Organization**: LLM-course
 
-## Model Details
-
+### Model Details
 - **Architecture**: Chess Transformer (GPT-style)
-- **Vocab size**: 72
+- **Vocab size**: 156
 - **Embedding dim**: 128
 - **Layers**: 6
 - **Heads**: 8

config.json

@@ -2,6 +2,10 @@
   "architectures": [
     "ChessForCausalLM"
   ],
+  "auto_map": {
+    "AutoConfig": "model.ChessConfig",
+    "AutoModelForCausalLM": "model.ChessForCausalLM"
+  },
   "bos_token_id": 1,
   "dropout": 0.1,
   "dtype": "float32",
@@ -11,10 +15,10 @@
   "n_ctx": 256,
   "n_embd": 128,
   "n_head": 8,
-  "n_inner": 360,
+  "n_inner": 354,
   "n_layer": 6,
   "pad_token_id": 0,
   "tie_weights": true,
-  "transformers_version": "4.57.5",
-  "vocab_size": 72
+  "transformers_version": "4.57.6",
+  "vocab_size": 156
 }

model.py

@@ -0,0 +1,437 @@
+"""
+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)
+        
+        # Declare tied weights for proper serialization
+        if config.tie_weights:
+            self._tied_weights_keys = ["lm_head.weight"]
+        
+        # Initialize weights
+        self.post_init()
+        
+        # Tie weights if configured
+        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):
+        # Use HF helper to tie or clone depending on config
+        if getattr(self.config, "tie_weights", False) or getattr(self.config, "tie_word_embeddings", False):
+            self._tie_or_clone_weights(self.lm_head, self.wte)
+    
+    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 = 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

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:96e89751fab2b60548bd62e6d1b132137e69909a1ad95b159be791149a048c76
-size 3996400
+oid sha256:71034cb4dec8ca9f0d453d40ce48f28ebecb8e4e2944f531df2e850b5fa61354
+size 4002400

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

@@ -33,12 +33,18 @@
       "special": true
     }
   },
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenizer_decomposed.ChessDecomposedTokenizer",
+      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",
+  "tokenizer_class": "ChessDecomposedTokenizer",
   "unk_token": "[UNK]"
 }

tokenizer_decomposed.py

@@ -0,0 +1,157 @@
+"""
+Decomposed Chess Tokenizer.
+
+This tokenizer decomposes each move into 3-4 tokens:
+- color+piece token (e.g., "WP", "BN")
+- from-square token with suffix "_f" (e.g., "e2_f")
+- to-square token with suffix "_t" (e.g., "e4_t")
+- optional promotion token (one of "q", "r", "b", "n")
+
+This avoids UNKs for rare moves and makes legality learning easier because the model
+always emits explicit squares.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+from typing import Dict, List, Optional
+
+from transformers import PreTrainedTokenizer
+
+
+class ChessDecomposedTokenizer(PreTrainedTokenizer):
+    model_input_names = ["input_ids", "attention_mask"]
+    vocab_files_names = {"vocab_file": "vocab.json"}
+
+    PAD_TOKEN = "[PAD]"
+    BOS_TOKEN = "[BOS]"
+    EOS_TOKEN = "[EOS]"
+    UNK_TOKEN = "[UNK]"
+
+    _MOVE_RE = re.compile(r"^[WB][PNBRQK][a-h][1-8][a-h][1-8].*$")
+
+    def __init__(
+        self,
+        vocab_file: Optional[str] = None,
+        vocab: Optional[Dict[str, int]] = None,
+        **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)
+
+        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:
+            self._vocab = self._create_full_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,
+        )
+
+    @staticmethod
+    def _create_full_vocab() -> Dict[str, int]:
+        special_tokens = [
+            ChessDecomposedTokenizer.PAD_TOKEN,
+            ChessDecomposedTokenizer.BOS_TOKEN,
+            ChessDecomposedTokenizer.EOS_TOKEN,
+            ChessDecomposedTokenizer.UNK_TOKEN,
+        ]
+
+        pieces = ["P", "N", "B", "R", "Q", "K"]
+        colors = ["W", "B"]
+        piece_tokens = [f"{c}{p}" for c in colors for p in pieces]
+
+        files = "abcdefgh"
+        ranks = "12345678"
+        squares = [f"{f}{r}" for f in files for r in ranks]
+        from_tokens = [f"{sq}_f" for sq in squares]
+        to_tokens = [f"{sq}_t" for sq in squares]
+
+        promo_tokens = ["q", "r", "b", "n"]
+
+        tokens = special_tokens + piece_tokens + from_tokens + to_tokens + promo_tokens
+        return {tok: idx for idx, tok in enumerate(tokens)}
+
+    @property
+    def vocab_size(self) -> int:
+        return len(self._vocab)
+
+    def get_vocab(self) -> Dict[str, int]:
+        return dict(self._vocab)
+
+    def _tokenize(self, text: str) -> List[str]:
+        raw = text.strip()
+        if not raw:
+            return []
+
+        parts = raw.split()
+        out: List[str] = []
+
+        for part in parts:
+            if part in {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}:
+                out.append(part)
+                continue
+
+            if not self._MOVE_RE.match(part):
+                out.append(self.UNK_TOKEN)
+                continue
+
+            color = part[0]
+            piece = part[1]
+            from_sq = part[2:4]
+            to_sq = part[4:6]
+            out.append(f"{color}{piece}")
+            out.append(f"{from_sq}_f")
+            out.append(f"{to_sq}_t")
+
+            if "=" in part:
+                promo_idx = part.find("=")
+                if promo_idx != -1 and promo_idx + 1 < len(part):
+                    promo = part[promo_idx + 1].lower()
+                    if promo in {"q", "r", "b", "n"}:
+                        out.append(promo)
+
+        return out
+
+    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 convert_tokens_to_string(self, tokens: List[str]) -> str:
+        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:
+        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,)
+

vocab.json

@@ -4,71 +4,155 @@
   "[EOS]": 2,
   "[UNK]": 3,
   "a1": 4,
-  "a2": 5,
-  "a3": 6,
-  "a4": 7,
-  "a5": 8,
-  "a6": 9,
-  "a7": 10,
-  "a8": 11,
-  "b1": 12,
+  "b1": 5,
+  "c1": 6,
+  "d1": 7,
+  "e1": 8,
+  "f1": 9,
+  "g1": 10,
+  "h1": 11,
+  "a2": 12,
   "b2": 13,
-  "b3": 14,
-  "b4": 15,
-  "b5": 16,
-  "b6": 17,
-  "b7": 18,
-  "b8": 19,
-  "c1": 20,
-  "c2": 21,
+  "c2": 14,
+  "d2": 15,
+  "e2": 16,
+  "f2": 17,
+  "g2": 18,
+  "h2": 19,
+  "a3": 20,
+  "b3": 21,
   "c3": 22,
-  "c4": 23,
-  "c5": 24,
-  "c6": 25,
-  "c7": 26,
-  "c8": 27,
-  "d1": 28,
-  "d2": 29,
-  "d3": 30,
+  "d3": 23,
+  "e3": 24,
+  "f3": 25,
+  "g3": 26,
+  "h3": 27,
+  "a4": 28,
+  "b4": 29,
+  "c4": 30,
   "d4": 31,
-  "d5": 32,
-  "d6": 33,
-  "d7": 34,
-  "d8": 35,
-  "e1": 36,
-  "e2": 37,
-  "e3": 38,
-  "e4": 39,
+  "e4": 32,
+  "f4": 33,
+  "g4": 34,
+  "h4": 35,
+  "a5": 36,
+  "b5": 37,
+  "c5": 38,
+  "d5": 39,
   "e5": 40,
-  "e6": 41,
-  "e7": 42,
-  "e8": 43,
-  "f1": 44,
-  "f2": 45,
-  "f3": 46,
-  "f4": 47,
-  "f5": 48,
+  "f5": 41,
+  "g5": 42,
+  "h5": 43,
+  "a6": 44,
+  "b6": 45,
+  "c6": 46,
+  "d6": 47,
+  "e6": 48,
   "f6": 49,
-  "f7": 50,
-  "f8": 51,
-  "g1": 52,
-  "g2": 53,
-  "g3": 54,
-  "g4": 55,
-  "g5": 56,
-  "g6": 57,
+  "g6": 50,
+  "h6": 51,
+  "a7": 52,
+  "b7": 53,
+  "c7": 54,
+  "d7": 55,
+  "e7": 56,
+  "f7": 57,
   "g7": 58,
-  "g8": 59,
-  "h1": 60,
-  "h2": 61,
-  "h3": 62,
-  "h4": 63,
-  "h5": 64,
-  "h6": 65,
-  "h7": 66,
+  "h7": 59,
+  "a8": 60,
+  "b8": 61,
+  "c8": 62,
+  "d8": 63,
+  "e8": 64,
+  "f8": 65,
+  "g8": 66,
   "h8": 67,
-  "q": 68,
-  "r": 69,
-  "b": 70,
-  "n": 71
+  "V:N:1": 68,
+  "V:N:2": 69,
+  "V:N:3": 70,
+  "V:N:4": 71,
+  "V:N:5": 72,
+  "V:N:6": 73,
+  "V:N:7": 74,
+  "V:S:1": 75,
+  "V:S:2": 76,
+  "V:S:3": 77,
+  "V:S:4": 78,
+  "V:S:5": 79,
+  "V:S:6": 80,
+  "V:S:7": 81,
+  "V:E:1": 82,
+  "V:E:2": 83,
+  "V:E:3": 84,
+  "V:E:4": 85,
+  "V:E:5": 86,
+  "V:E:6": 87,
+  "V:E:7": 88,
+  "V:W:1": 89,
+  "V:W:2": 90,
+  "V:W:3": 91,
+  "V:W:4": 92,
+  "V:W:5": 93,
+  "V:W:6": 94,
+  "V:W:7": 95,
+  "V:NE:1": 96,
+  "V:NE:2": 97,
+  "V:NE:3": 98,
+  "V:NE:4": 99,
+  "V:NE:5": 100,
+  "V:NE:6": 101,
+  "V:NE:7": 102,
+  "V:NW:1": 103,
+  "V:NW:2": 104,
+  "V:NW:3": 105,
+  "V:NW:4": 106,
+  "V:NW:5": 107,
+  "V:NW:6": 108,
+  "V:NW:7": 109,
+  "V:SE:1": 110,
+  "V:SE:2": 111,
+  "V:SE:3": 112,
+  "V:SE:4": 113,
+  "V:SE:5": 114,
+  "V:SE:6": 115,
+  "V:SE:7": 116,
+  "V:SW:1": 117,
+  "V:SW:2": 118,
+  "V:SW:3": 119,
+  "V:SW:4": 120,
+  "V:SW:5": 121,
+  "V:SW:6": 122,
+  "V:SW:7": 123,
+  "V:KN:NNE": 124,
+  "V:KN:ENE": 125,
+  "V:KN:ESE": 126,
+  "V:KN:SSE": 127,
+  "V:KN:SSW": 128,
+  "V:KN:WSW": 129,
+  "V:KN:WNW": 130,
+  "V:KN:NNW": 131,
+  "V:N:1:=Q": 132,
+  "V:N:1:=R": 133,
+  "V:N:1:=B": 134,
+  "V:N:1:=N": 135,
+  "V:NE:1:=Q": 136,
+  "V:NE:1:=R": 137,
+  "V:NE:1:=B": 138,
+  "V:NE:1:=N": 139,
+  "V:NW:1:=Q": 140,
+  "V:NW:1:=R": 141,
+  "V:NW:1:=B": 142,
+  "V:NW:1:=N": 143,
+  "V:S:1:=Q": 144,
+  "V:S:1:=R": 145,
+  "V:S:1:=B": 146,
+  "V:S:1:=N": 147,
+  "V:SE:1:=Q": 148,
+  "V:SE:1:=R": 149,
+  "V:SE:1:=B": 150,
+  "V:SE:1:=N": 151,
+  "V:SW:1:=Q": 152,
+  "V:SW:1:=R": 153,
+  "V:SW:1:=B": 154,
+  "V:SW:1:=N": 155
 }