Chess Challenge submission by MarcosElFlamenco

README.md

@@ -0,0 +1,22 @@
+---
+library_name: transformers
+tags:
+- chess
+- llm-course
+- chess-challenge
+license: mit
+---
+
+## Chess model submitted to the LLM Course Chess Challenge.
+
+### Submission Info
+- **Submitted by**: [MarcosElFlamenco](https://huggingface.co/MarcosElFlamenco)
+- **Parameters**: 964,534
+- **Organization**: LLM-course
+
+### Model Details
+- **Architecture**: Chess Transformer (GPT-style)
+- **Vocab size**: 82
+- **Embedding dim**: 124
+- **Layers**: 7
+- **Heads**: 4

config.json

@@ -0,0 +1,26 @@
+{
+  "architectures": [
+    "ChessForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "model.ChessConfig",
+    "AutoModelForCausalLM": "model.ChessForCausalLM"
+  },
+  "bos_token_id": 1,
+  "dropout": 0.1,
+  "dtype": "float32",
+  "eos_token_id": 2,
+  "group_size": 4,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "chess_transformer",
+  "n_ctx": 256,
+  "n_embd": 124,
+  "n_head": 4,
+  "n_inner": 372,
+  "n_layer": 7,
+  "pad_token_id": 0,
+  "rms_Norm": true,
+  "tie_weights": true,
+  "transformers_version": "4.57.5",
+  "vocab_size": 83
+}

model.py

@@ -0,0 +1,531 @@
+"""
+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.
+        rms_Norm: Whether to use RMSNorm instead of LayerNorm.
+    
+    """
+    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,
+        group_size: Optional[int] = None,
+        dropout: float = 0.1,
+        layer_norm_epsilon: float = 1e-5,
+        tie_weights: bool = True,
+        rms_Norm: bool = False,
+        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.group_size = group_size
+        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
+        self.rms_Norm = rms_Norm
+        # Inform HF base class about tying behavior
+        self.tie_word_embeddings = bool(tie_weights)
+
+
+class GroupedQueryAttention(nn.Module):
+
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+
+        assert config.n_head % config.group_size == 0, "n_head must be divisible by group_size"
+        print(f"Using Grouped Query Attention with group_size={config.group_size}")        
+        self.n_head = config.n_head          # Total Query heads
+        self.group_size = config.group_size
+        self.n_kv_head = self.n_head // config.group_size  # Number of KV heads
+        
+        self.n_embd = config.n_embd
+        self.head_dim = config.n_embd // config.n_head
+        
+        # Q projection stays the same, but K and V projections are smaller
+        # Total output: n_embd (for Q) + 2 * (n_kv_head * head_dim) (for K and V)
+        self.q_proj = nn.Linear(config.n_embd, config.n_embd)
+        self.kv_proj = nn.Linear(config.n_embd, 2 * self.n_kv_head * self.head_dim)
+        
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd)
+        self.dropout = nn.Dropout(config.dropout)
+        
+        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()
+        
+        # 1. Project Q, K, V
+        q = self.q_proj(x)  # (B, T, n_head * head_dim)
+        kv = self.kv_proj(x) # (B, T, 2 * n_kv_head * head_dim)
+        k, v = kv.split(self.n_kv_head * self.head_dim, dim=2)
+        
+        # 2. Reshape Q normally
+        q = q.view(batch_size, seq_len, self.n_head, self.head_dim).transpose(1, 2)
+        
+        # 3. Reshape K, V and REPEAT them to match Q
+        k = k.view(batch_size, seq_len, self.n_kv_head, self.head_dim).transpose(1, 2)
+        v = v.view(batch_size, seq_len, self.n_kv_head, self.head_dim).transpose(1, 2)
+        
+        # Repeat KV heads 'group_size' times to match n_head
+        # We use .repeat_interleave to ensure head 0 of KV is used by the first 'group_size' Q heads
+        k = k.repeat_interleave(self.group_size, dim=1) # (B, n_head, T, head_dim)
+        v = v.repeat_interleave(self.group_size, dim=1) # (B, n_head, T, head_dim)
+        
+        # 4. Standard Scaled Dot-Product Attention
+        attn_weights = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
+        
+        causal_mask = self.bias[:, :, :seq_len, :seq_len]
+        attn_weights = attn_weights.masked_fill(causal_mask == 0, float("-inf"))
+        
+        if attention_mask is not None:
+            attn_weights = attn_weights.masked_fill(attention_mask.unsqueeze(1).unsqueeze(2) == 0, float("-inf"))
+        
+        attn_weights = F.softmax(attn_weights, dim=-1)
+        attn_output = torch.matmul(self.dropout(attn_weights), v)
+        
+        # 5. Recombine
+        attn_output = attn_output.transpose(1, 2).contiguous().view(batch_size, seq_len, self.n_embd)
+        return self.c_proj(attn_output)
+    
+
+
+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})"
+        
+        print(f"Using Regular Attention with group_size={config.group_size}")        
+        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,group_size: int = None):
+        super().__init__()
+#nn.modules.normalization.RMSNorm
+        if config.rms_Norm == True:
+            print(f"using RMSNorm")
+            self.ln_1 = nn.RMSNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        else:
+            print(f"using LayerNorm")
+            self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+
+        if config.group_size is not None:
+            self.attn = GroupedQueryAttention(config)
+        else:
+            self.attn = MultiHeadAttention(config)
+        if config.rms_Norm == True:
+            self.ln_2 = nn.RMSNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        else:
+            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_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.
+        
+        Applies structural constraints enforcing:
+        ColoredPiece [SOURCE] source [DEST] dest [modifiers]*
+        
+        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, :].clone() / temperature
+        
+        # Apply structural constraints (hardcoded to ChessTokenizer structure)
+        from src.tokenizer import ChessLogitsProcessor
+        processor = ChessLogitsProcessor()
+        logits = processor.constrain_logits(input_ids, logits)
+        
+        # Apply top-k filtering
+        if top_k is not None:
+            indices_to_remove = logits < torch.topk(logits, min(top_k, logits.size(-1)))[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:c22bdb64b6359302e943b73365d3ad45efc1b513c1fad1a5cc2937de7866676a
+size 3865736

special_tokens_map.json

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

tokenizer.py

@@ -0,0 +1,906 @@
+"""
+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 FrequencyChessTokenizer(PreTrainedTokenizer):
+    """
+    A frequency-based 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.
+    
+    Only includes moves that appear at least `min_frequency` times in the dataset.
+    Rare moves become [UNK] tokens.
+    
+    Example:
+        >>> tokenizer = FrequencyChessTokenizer()
+        >>> 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,
+    ) -> "FrequencyChessTokenizer":
+        """
+        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 FrequencyChessTokenizer 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,
+    ) -> "FrequencyChessTokenizer":
+        """
+        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 FrequencyChessTokenizer 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)
+
+
+class ChessTokenizer(FrequencyChessTokenizer):
+    """
+    A compositional tokenizer for chess moves using split color/piece tokens.
+    
+    This tokenizer breaks each move into 6 core components with explicit structure:
+    1. Color: W or B (makes turn information explicit!)
+    2. Piece: P, N, B, R, Q, K
+    3. SOURCE marker: [SOURCE]
+    4. Source square: a1, a2, ..., h8
+    5. DEST marker: [DEST]
+    6. Destination square: a1, a2, ..., h8
+    
+    Optional modifier tokens for captures, checks, checkmate, and castling.
+    
+    Example:
+        >>> tokenizer = ChessTokenizer()
+        >>> tokenizer.encode("WPe2e4 BPe7e5")
+        [1, W_id, P_id, SRC_id, e2_id, DST_id, e4_id, B_id, P_id, SRC_id, e7_id, DST_id, e5_id, 2]
+    
+    Vocabulary:
+    - Colors (2): W, B [makes turn alternation explicit]
+    - Pieces (6): P, N, B, R, Q, K
+    - Position markers (2): [SOURCE], [DEST]
+    - Squares (64): a1-h8
+    - Modifiers (5): [CAPTURE], [CHECK], [CHECKMATE], [CASTLING_KS], [CASTLING_QS]
+    - Special (4): [PAD], [BOS], [EOS], [UNK]
+    Total: ~83 tokens (deterministic, 4 fewer than before)
+    
+    Key advantage: Color is now EXPLICIT, making turn alternation obvious to the model!
+    """
+    
+    # Color tokens (split for explicit turn information)
+    COLORS = ['W', 'B']
+    
+    # Piece tokens
+    PIECES = ['P', 'N', 'B', 'R', 'Q', 'K']
+    
+    # Position markers
+    POSITION_MARKERS = ['[SOURCE]', '[DEST]']
+    
+    # Board squares (standard chess notation)
+    SQUARES = [f"{file}{rank}" for rank in range(1, 9) for file in "abcdefgh"]
+    
+    # Move modifiers
+    MODIFIERS = ['[CAPTURE]', '[CHECK]', '[CHECKMATE]', '[CASTLING_KS]', '[CASTLING_QS]']
+    
+    def __init__(self, **kwargs):
+        """
+        Initialize the compositional chess tokenizer.
+        
+        Vocabulary is built deterministically from pieces and squares.
+        No vocab_file or dataset scanning needed.
+        """
+        # Remove vocab-related kwargs to avoid conflicts
+        kwargs.pop("vocab_file", None)
+        kwargs.pop("vocab", None)
+        
+        # Build deterministic vocabulary
+        vocab = self._build_deterministic_vocab()
+        
+        # Initialize parent with the built vocab
+        super().__init__(vocab=vocab, **kwargs)
+    
+    @property
+    def vocab_size(self) -> int:
+        """
+        Return the vocabulary size.
+        
+        Tokens: [PAD]=0, [BOS]=1, [EOS]=2, [UNK]=3, W=4, B=5, P-K=6-11, 
+                 [SOURCE]=12, [DEST]=13, squares=14-77, modifiers=78-82
+        
+        Total: 83 tokens (indices 0-82)
+        """
+        return 4 + 2 + 6 + 2 + 64 + 5  # special + colors + pieces + markers + squares + modifiers
+    
+    def _build_deterministic_vocab(self) -> Dict[str, int]:
+        """
+        Build vocabulary deterministically from colored pieces, squares, and modifiers.
+        
+        Returns:
+            Dictionary mapping token strings to IDs.
+        """
+        vocab = {}
+        idx = 0
+        
+        # Special tokens first (matching parent class order)
+        special_tokens = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]
+        for token in special_tokens:
+            vocab[token] = idx
+            idx += 1
+        
+        # Color tokens (W, B)
+        for color in self.COLORS:
+            vocab[color] = idx
+            idx += 1
+        
+        # Piece tokens (P, N, B, R, Q, K)
+        for piece in self.PIECES:
+            vocab[piece] = idx
+            idx += 1
+        
+        # Position marker tokens
+        for marker in self.POSITION_MARKERS:
+            vocab[marker] = idx
+            idx += 1
+        
+        # Square tokens
+        for square in self.SQUARES:
+            vocab[square] = idx
+            idx += 1
+        
+        # Modifier tokens
+        for modifier in self.MODIFIERS:
+            vocab[modifier] = idx
+            idx += 1
+        
+        return vocab
+    
+    def _parse_move(self, move_str: str) -> Dict:
+        """
+        Parse a move string in extended UCI notation.
+        
+        Args:
+            move_str: Move string like "WPe2e4" or "BNg8f6(x)" or "We1g1(o)"
+        
+        Returns:
+            Dictionary with keys: piece, color, src, dest, modifiers
+        """
+        import re
+        
+        # Pattern: [WB][PNBRQK]<square><square>(<modifiers>)
+        pattern = r'([WB])([PNBRQK])([a-h][1-8])([a-h][1-8])((?:\([^)]*\))?)'
+        match = re.match(pattern, move_str.strip())
+        
+        if not match:
+            raise ValueError(f"Invalid move format: {move_str}")
+        
+        color, piece, src, dest, modifier_str = match.groups()
+        
+        # Parse modifiers
+        modifiers = []
+        if modifier_str:
+            # Remove parentheses and split by lowercase letters/symbols
+            mod_content = modifier_str.strip('()')
+            
+            if 'x' in mod_content:
+                modifiers.append('[CAPTURE]')
+            if '+*' in mod_content:
+                modifiers.append('[CHECKMATE]')
+            elif '+' in mod_content:
+                modifiers.append('[CHECK]')
+            if 'o' in mod_content or 'O' in mod_content:
+                # Determine kingside vs queenside based on destination
+                if dest == 'g1' or dest == 'g8':
+                    modifiers.append('[CASTLING_KS]')
+                elif dest == 'c1' or dest == 'c8':
+                    modifiers.append('[CASTLING_QS]')
+        
+        return {
+            'piece': piece,
+            'color': color,
+            'src': src,
+            'dest': dest,
+            'modifiers': modifiers,
+        }
+    
+    def _tokenize(self, text: str) -> List[str]:
+        """
+        Tokenize a string of moves into component tokens with positional markers.
+        
+        Each move becomes: [ColoredPiece, [SOURCE], source, [DEST], dest, *modifiers]
+        
+        Args:
+            text: String of space-separated moves (e.g., "WPe2e4 BPe7e5")
+        
+        Returns:
+            List of component tokens with structure markers.
+        """
+        move_strings = text.strip().split()
+        tokens = []
+        
+        for move_str in move_strings:
+            parsed = self._parse_move(move_str)
+            
+            # Add color and piece as SEPARATE tokens (now explicit!)
+            tokens.append(parsed['color'])  # W or B
+            tokens.append(parsed['piece'])  # P, N, B, R, Q, K
+            
+            # Add positional markers and squares
+            tokens.append('[SOURCE]')
+            tokens.append(parsed['src'])
+            tokens.append('[DEST]')
+            tokens.append(parsed['dest'])
+            
+            # Add modifier tokens if any
+            tokens.extend(parsed['modifiers'])
+        
+        return tokens
+    
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        """
+        Reconstruct moves from component tokens with positional markers.
+        
+        Expects structure: Color, Piece, [SOURCE], source, [DEST], dest, *modifiers
+        
+        Args:
+            tokens: List of component tokens
+        
+        Returns:
+            Space-separated move string.
+        """
+        moves = []
+        token_idx = 0
+        
+        while token_idx < len(tokens):
+            token = tokens[token_idx]
+            
+            # Skip special tokens
+            special = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}
+            if token in special:
+                token_idx += 1
+                continue
+            
+            # Expect: Color token (W or B)
+            if token not in self.COLORS:
+                break
+            
+            color = token
+            
+            # Expect: Piece token (P, N, B, R, Q, K)
+            if token_idx + 1 >= len(tokens) or tokens[token_idx + 1] not in self.PIECES:
+                break
+            
+            piece = tokens[token_idx + 1]
+            colored_piece = color + piece
+            
+            # Expect: [SOURCE] marker
+            if token_idx + 2 >= len(tokens) or tokens[token_idx + 2] != '[SOURCE]':
+                break
+            
+            # Expect: source square
+            if token_idx + 3 >= len(tokens):
+                break
+            src = tokens[token_idx + 3]
+            if src not in self.SQUARES:
+                break
+            
+            # Expect: [DEST] marker
+            if token_idx + 4 >= len(tokens) or tokens[token_idx + 4] != '[DEST]':
+                break
+            
+            # Expect: dest square
+            if token_idx + 5 >= len(tokens):
+                break
+            dest = tokens[token_idx + 5]
+            if dest not in self.SQUARES:
+                break
+            
+            # Build move string
+            move_str = f"{color}{piece}{src}{dest}"
+            
+            # Collect modifiers (next tokens until we hit another color token or end)
+            token_idx += 6
+            modifiers_list = []
+            
+            while token_idx < len(tokens) and tokens[token_idx] in self.MODIFIERS:
+                modifier = tokens[token_idx]
+                modifiers_list.append(modifier)
+                token_idx += 1
+            
+            # Append modifier suffixes
+            if modifiers_list:
+                modifier_str = ""
+                if '[CAPTURE]' in modifiers_list:
+                    modifier_str += "x"
+                if '[CHECKMATE]' in modifiers_list:
+                    modifier_str += "+*"
+                elif '[CHECK]' in modifiers_list:
+                    modifier_str += "+"
+                if '[CASTLING_KS]' in modifiers_list:
+                    modifier_str += "o"
+                elif '[CASTLING_QS]' in modifiers_list:
+                    modifier_str += "o"
+                
+                move_str += f"({modifier_str})"
+            
+            moves.append(move_str)
+        
+        return " ".join(moves)
+    
+    def decode(self, token_ids, skip_special_tokens=False, **kwargs):
+        """
+        Decode token IDs back to string representation.
+        
+        Properly handles individual tokens by converting each ID to its token string.
+        For single tokens or incomplete move sequences, returns the raw token strings.
+        For complete move sequences, reconstructs the move format.
+        
+        Args:
+            token_ids: List or tensor of token IDs
+            skip_special_tokens: Whether to skip special tokens in output
+            **kwargs: Additional arguments (for compatibility)
+        
+        Returns:
+            String representation of the tokens
+        """
+        # Convert tensor to list if needed
+        if hasattr(token_ids, 'tolist'):
+            token_ids = token_ids.tolist()
+        
+        # Handle 2D tensor/list (batch)
+        if isinstance(token_ids, list) and len(token_ids) > 0 and isinstance(token_ids[0], list):
+            return [self.decode(ids, skip_special_tokens=skip_special_tokens) for ids in token_ids]
+        
+        # Convert IDs to tokens
+        tokens = []
+        for token_id in token_ids:
+            if isinstance(token_id, int):
+                token = self._convert_id_to_token(token_id)
+            else:
+                token = str(token_id)
+            
+            tokens.append(token)
+        
+        # Try to reconstruct moves from tokens
+        # If successful, return the reconstructed moves
+        reconstructed = self._try_reconstruct_moves(tokens, skip_special_tokens)
+        if reconstructed is not None:
+            return reconstructed
+        
+        # Fallback: return tokens joined with spaces, filtering special tokens if requested
+        if skip_special_tokens:
+            special = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}
+            tokens = [t for t in tokens if t not in special]
+        
+        return " ".join(tokens)
+    
+    def _try_reconstruct_moves(self, tokens: List[str], skip_special_tokens: bool = False) -> Optional[str]:
+        """
+        Try to reconstruct complete moves from tokens.
+        
+        Returns the reconstructed move string if tokens form valid move(s),
+        None if tokens don't form a complete move structure.
+        
+        Args:
+            tokens: List of token strings
+            skip_special_tokens: Whether to skip special tokens
+        
+        Returns:
+            Reconstructed move string or None
+        """
+        moves = []
+        token_idx = 0
+        found_moves = False
+        
+        while token_idx < len(tokens):
+            token = tokens[token_idx]
+            
+            # Skip special tokens
+            special = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}
+            if token in special:
+                token_idx += 1
+                continue
+            
+            # Check if this starts a move (color token)
+            if token not in self.COLORS:
+                # No more complete moves
+                break
+            
+            color = token
+            
+            # Need at least 6 more tokens for a complete move
+            if token_idx + 5 >= len(tokens):
+                break
+            
+            # Expect: Piece token (P, N, B, R, Q, K)
+            if tokens[token_idx + 1] not in self.PIECES:
+                break
+            
+            piece = tokens[token_idx + 1]
+            
+            # Expect: [SOURCE] marker
+            if tokens[token_idx + 2] != '[SOURCE]':
+                break
+            
+            # Expect: source square
+            src = tokens[token_idx + 3]
+            if src not in self.SQUARES:
+                break
+            
+            # Expect: [DEST] marker
+            if tokens[token_idx + 4] != '[DEST]':
+                break
+            
+            # Expect: dest square
+            dest = tokens[token_idx + 5]
+            if dest not in self.SQUARES:
+                break
+            
+            # Build move string
+            move_str = f"{color}{piece}{src}{dest}"
+            
+            # Collect modifiers
+            token_idx += 6
+            modifiers_list = []
+            
+            while token_idx < len(tokens) and tokens[token_idx] in self.MODIFIERS:
+                modifiers_list.append(tokens[token_idx])
+                token_idx += 1
+            
+            # Append modifier suffixes
+            if modifiers_list:
+                modifier_str = ""
+                if '[CAPTURE]' in modifiers_list:
+                    modifier_str += "x"
+                if '[CHECKMATE]' in modifiers_list:
+                    modifier_str += "+*"
+                elif '[CHECK]' in modifiers_list:
+                    modifier_str += "+"
+                if '[CASTLING_KS]' in modifiers_list:
+                    modifier_str += "o"
+                elif '[CASTLING_QS]' in modifiers_list:
+                    modifier_str += "o"
+                
+                move_str += f"({modifier_str})"
+            
+            moves.append(move_str)
+            found_moves = True
+        
+        if found_moves:
+            return " ".join(moves)
+        
+        return None
+
+
+class ChessLogitsProcessor:
+    """
+    Logits processor for enforcing chess move structure during generation.
+    
+    Enforces the token sequence pattern:
+    Color Piece [SOURCE] source [DEST] dest [modifiers]*
+    
+    Uses a state machine with 7 states:
+    - State 0: Expect color (W, B)
+    - State 1: Expect piece (P, N, B, R, Q, K)
+    - State 2: Expect [SOURCE] marker
+    - State 3: Expect source square (a1-h8)
+    - State 4: Expect [DEST] marker
+    - State 5: Expect dest square (a1-h8)
+    - State 6: Expect modifiers or next color token
+    
+    Token structure is hardcoded to match ChessTokenizer:
+    - Colors: W, B (EXPLICIT for turn alternation)
+    - Pieces: P, N, B, R, Q, K
+    - Position markers: [SOURCE], [DEST]
+    - Squares: a1-h8 (64 total)
+    - Modifiers: [CAPTURE], [CHECK], [CHECKMATE], [CASTLING_KS], [CASTLING_QS]
+    """
+    
+    # Token vocabulary indices (hardcoded to match ChessTokenizer vocab order)
+    # Special tokens: [PAD]=0, [BOS]=1, [EOS]=2, [UNK]=3
+    # Colors (4-5)
+    COLOR_IDS = {'W': 4, 'B': 5}
+    # Pieces (6-11)
+    PIECE_IDS = {'P': 6, 'N': 7, 'B': 8, 'R': 9, 'Q': 10, 'K': 11}
+    # Position markers (12-13)
+    POSITION_MARKER_IDS = {'[SOURCE]': 12, '[DEST]': 13}
+    # Squares (14-77): a1=14, a2=15, ..., h8=77
+    SQUARE_IDS = {f"{file}{rank}": 14 + (rank - 1) * 8 + ord(file) - ord('a')
+                  for rank in range(1, 9) for file in "abcdefgh"}
+    # Modifiers (78-82)
+    MODIFIER_IDS = {
+        '[CAPTURE]': 78, '[CHECK]': 79, '[CHECKMATE]': 80,
+        '[CASTLING_KS]': 81, '[CASTLING_QS]': 82
+    }
+    
+    def __init__(self):
+        """
+        Initialize the logits processor with hardcoded ChessTokenizer structure.
+        """
+        import torch
+        self.torch = torch
+        
+        # Convert to sets for membership testing
+        self.color_ids = set(self.COLOR_IDS.values())
+        self.piece_ids = set(self.PIECE_IDS.values())
+        self.square_ids = set(self.SQUARE_IDS.values())
+        self.modifier_ids = set(self.MODIFIER_IDS.values())
+    
+    def _get_state(self, input_ids):
+        """
+        Determine current state in move sequence based on recent tokens.
+        
+        Returns state (0-6) indicating what token type is expected next.
+        """
+        if input_ids.numel() == 0:
+            return 0  # Start: expect color
+        
+        # Get the sequence of tokens
+        seq = input_ids[0].tolist()
+        
+        # Work backwards to find the last color token (marks start of move)
+        last_move_idx = -1
+        for i in range(len(seq) - 1, -1, -1):
+            if seq[i] in self.color_ids:
+                last_move_idx = i
+                break
+        
+        if last_move_idx == -1:
+            return 0  # No color found, expect color
+        
+        # Count tokens since last color
+        tokens_since_color = len(seq) - 1 - last_move_idx
+        
+        # Pattern: Color, Piece, [SOURCE], source, [DEST], dest, ...modifiers
+        if tokens_since_color == 0:
+            return 1  # Expect piece after color
+        elif tokens_since_color == 1:
+            # Should have: color, piece
+            if seq[-1] in self.piece_ids:
+                return 2  # Expect [SOURCE]
+            else:
+                return 1  # Unexpected, reset
+        elif tokens_since_color == 2:
+            # Should have: color, piece, [SOURCE]
+            if (seq[-2] in self.piece_ids and 
+                seq[-1] in [self.POSITION_MARKER_IDS['[SOURCE]']]):
+                return 3  # Expect source square
+            else:
+                return 1  # Reset
+        elif tokens_since_color == 3:
+            # Should have: color, piece, [SOURCE], source
+            if (seq[-3] in self.piece_ids and
+                seq[-2] in [self.POSITION_MARKER_IDS['[SOURCE]']] and
+                seq[-1] in self.square_ids):
+                return 4  # Expect [DEST]
+            else:
+                return 1  # Reset
+        elif tokens_since_color == 4:
+            # Should have: color, piece, [SOURCE], source, [DEST]
+            if (seq[-2] in self.square_ids and
+                seq[-1] in [self.POSITION_MARKER_IDS['[DEST]']]):
+                return 5  # Expect dest square
+            else:
+                return 1  # Reset
+        elif tokens_since_color == 5:
+            # Should have: color, piece, [SOURCE], source, [DEST], dest
+            if seq[-1] in self.square_ids:
+                return 6  # Expect modifiers or next color (move complete)
+            else:
+                return 1  # Reset
+        else:
+            # tokens_since_color >= 6: We're in modifiers or expecting next move
+            # If last token is a modifier, still expect more modifiers or next color
+            # If last token is not a modifier, we should expect next color
+            if seq[-1] not in self.modifier_ids:
+                return 0  # Expect next move (next color)
+            else:
+                return 6  # Could be more modifiers or next color
+    
+    def constrain_logits(self, input_ids, logits):
+        """
+        Mask invalid tokens in logits based on move structure.
+        
+        Sets logits to -inf for tokens that violate move structure.
+        
+        Args:
+            input_ids: Model input token IDs of shape (batch_size, seq_len)
+            logits: Model output logits of shape (batch_size, vocab_size)
+        
+        Returns:
+            Modified logits with invalid tokens masked to -inf
+        """
+        state = self._get_state(input_ids)
+        
+        # Create a mask for valid tokens (all ones initially)
+        valid_mask = self.torch.ones(logits.shape[-1], dtype=self.torch.bool)
+        valid_mask[:] = False  # Start by forbidding all
+        
+        # Allow tokens based on current state
+        if state == 0:
+            # Expect color (W or B)
+            for color_id in self.color_ids:
+                valid_mask[color_id] = True
+        
+        elif state == 1:
+            # Expect piece (P, N, B, R, Q, K)
+            for piece_id in self.piece_ids:
+                valid_mask[piece_id] = True
+        
+        elif state == 2:
+            # Expect [SOURCE]
+            valid_mask[self.POSITION_MARKER_IDS['[SOURCE]']] = True
+        
+        elif state == 3:
+            # Expect source square
+            for square_id in self.square_ids:
+                valid_mask[square_id] = True
+        
+        elif state == 4:
+            # Expect [DEST]
+            valid_mask[self.POSITION_MARKER_IDS['[DEST]']] = True
+        
+        elif state == 5:
+            # Expect dest square
+            for square_id in self.square_ids:
+                valid_mask[square_id] = True
+        
+        elif state == 6:
+            # Expect modifiers or next color token
+            # Allow: modifiers + colors + EOS
+            for modifier_id in self.modifier_ids:
+                valid_mask[modifier_id] = True
+            for color_id in self.color_ids:
+                valid_mask[color_id] = True
+            valid_mask[2] = True  # Allow EOS to end sequence
+        
+        # Apply mask
+        logits = logits.clone()
+        logits[0, ~valid_mask] = float('-inf')
+        
+        return logits
+

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_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": "ChessDecomposedTokenizer",
+  "unk_token": "[UNK]"
+}

tokenizer_decomposed.py

@@ -0,0 +1,156 @@
+"""
+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

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