Chess Challenge submission by Gusthavok

README.md

@@ -0,0 +1,31 @@
+---
+library_name: transformers
+tags:
+- chess
+- llm-course
+- chess-challenge
+license: mit
+---
+
+# chess_model_khg
+
+Chess model submitted to the LLM Course Chess Challenge.
+
+## Submission Info
+
+- **Submitted by**: [Gusthavok](https://huggingface.co/Gusthavok)
+- **Parameters**: 871,808
+- **Organization**: LLM-course
+
+## Usage
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model = AutoModelForCausalLM.from_pretrained("LLM-course/chess_model_khg", trust_remote_code=True)
+tokenizer = AutoTokenizer.from_pretrained("LLM-course/chess_model_khg", trust_remote_code=True)
+```
+
+## Evaluation
+
+This model is evaluated at the [Chess Challenge Arena](https://huggingface.co/spaces/LLM-course/Chess1MChallenge).

config.json

@@ -0,0 +1,24 @@
+{
+  "architectures": [
+    "ChessForCausalLM"
+  ],
+  "bos_token_id": 1,
+  "dropout": 0.1,
+  "dtype": "float32",
+  "eos_token_id": 2,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "chess_transformer",
+  "n_ctx": 256,
+  "n_embd": 128,
+  "n_head": 4,
+  "n_inner": 384,
+  "n_layer": 5,
+  "pad_token_id": 0,
+  "tie_weights": true,
+  "transformers_version": "4.56.2",
+  "vocab_size": 93,
+  "auto_map": {
+    "AutoConfig": "model.ChessConfig",
+    "AutoModelForCausalLM": "model.ChessForCausalLM"
+  }
+}

model.py

@@ -0,0 +1,643 @@
+"""
+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()
+        
+        # Structured generation support
+        self.vocab_masks = None
+        self.tokenizer_ref = None
+        self._auto_prepare_attempted = False
+        
+        # Enable constrained decoding during inference (disabled in training)
+        self.use_constrained_decoding = False
+    
+    def _setup_token_type_indices(self, device):
+        """Setup indices for different token types for constrained generation. Lazy initialization."""
+        if self._token_indices_initialized:
+            return
+        
+        # Color+piece tokens (WP, WN, WB, WR, WQ, WK, BP, BN, BB, BR, BQ, BK)
+        piece_tokens = ["WP", "WN", "WB", "WR", "WQ", "WK", "BP", "BN", "BB", "BR", "BQ", "BK"]
+        
+        # Position tokens (a1-h8)
+        files = ['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h']
+        ranks = ['1', '2', '3', '4', '5', '6', '7', '8']
+        square_tokens = [f + r for f in files for r in ranks]
+        
+        # Suffix tokens
+        suffix_tokens = ["(x)", "(+)", "(x+)", "(+*)", "(x+*)", "(o)", "(O)", "(xE)", "=Q", "=R", "=B", "=N"]
+        
+        # Special tokens
+        special_tokens = ["[PAD]", "[BOS]", "[EOS]", "[UNK]"]
+        
+        # Create vocab to index mapping (build a simple vocab dict)
+        # Assuming tokenizer uses this order: special tokens, pieces, squares, suffixes
+        vocab = {}
+        idx = 0
+        for token in special_tokens + piece_tokens + square_tokens + suffix_tokens:
+            vocab[token] = idx
+            idx += 1
+        
+        # Store indices on the correct device
+        self.piece_token_ids = torch.tensor([vocab[t] for t in piece_tokens if t in vocab], device=device)
+        self.square_token_ids = torch.tensor([vocab[t] for t in square_tokens if t in vocab], device=device)
+        self.suffix_token_ids = torch.tensor([vocab[t] for t in suffix_tokens if t in vocab], device=device)
+        self.special_token_ids = torch.tensor([vocab[t] for t in special_tokens if t in vocab], device=device)
+        
+        self._token_indices_initialized = True
+    
+    def get_constrained_logits_mask(self, input_ids: torch.LongTensor) -> torch.Tensor:
+        """
+        Create a mask for constrained decoding based on the move pattern.
+        
+        Pattern: [Piece] [Square] [Square] [Optional Suffix]
+        
+        Args:
+            input_ids: Input token IDs of shape (batch_size, seq_len)
+        
+        Returns:
+            Mask of shape (batch_size, vocab_size) where 1 = allowed, 0 = forbidden
+        """
+        batch_size, seq_len = input_ids.size()
+        device = input_ids.device
+        vocab_size = self.config.vocab_size
+        
+        
+        # Lazy initialization of token indices on the correct device
+        self._setup_token_type_indices(device)
+        
+        # Initialize mask (all tokens forbidden by default)
+        mask = torch.zeros(batch_size, vocab_size, device=device)
+        
+        # Token indices are already on the correct device
+        piece_ids = self.piece_token_ids
+        square_ids = self.square_token_ids
+        suffix_ids = self.suffix_token_ids
+        special_ids = self.special_token_ids
+        # Move indices to device
+        piece_ids = self.piece_token_ids.to(device)
+        square_ids = self.square_token_ids.to(device)
+        suffix_ids = self.suffix_token_ids.to(device)
+        special_ids = self.special_token_ids.to(device)
+        
+        for b in range(batch_size):
+            # Get recent tokens (look at last few to determine pattern position)
+            # Count backwards from the last piece token to determine position in move
+            recent_tokens = input_ids[b, max(0, seq_len-10):seq_len]
+            
+            # Find the last occurrence of a piece token
+            piece_mask = torch.isin(recent_tokens, piece_ids)
+            if piece_mask.any():
+                last_piece_idx = torch.where(piece_mask)[0][-1].item()
+                tokens_since_piece = len(recent_tokens) - last_piece_idx - 1
+                
+                if tokens_since_piece == 0:
+                    # Just saw piece → expect square (from_square)
+                    mask[b, square_ids] = 1
+                elif tokens_since_piece == 1:
+                    # Just saw from_square → expect square (to_square)
+                    mask[b, square_ids] = 1
+                elif tokens_since_piece == 2:
+                    # Just saw to_square → expect suffix OR new piece
+                    mask[b, suffix_ids] = 1
+                    mask[b, piece_ids] = 1
+                else:
+                    # After suffix or multiple tokens → expect new piece
+                    mask[b, piece_ids] = 1
+            else:
+                # No piece seen yet → expect piece (start of sequence/game)
+                mask[b, piece_ids] = 1
+                mask[b, special_ids] = 1  # Allow special tokens
+        
+        return mask
+
+    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 prepare_for_structured_generation(self, tokenizer):
+        """Prepare the model for structured generation with a tokenizer.
+        
+        This builds vocabulary masks and stores tokenizer reference for state analysis.
+        Call this before generation to enable structured move generation.
+        
+        Args:
+            tokenizer: ChessTokenizer instance with build_vocabulary_masks method.
+        """
+        self.tokenizer_ref = tokenizer
+        self.vocab_masks = tokenizer.build_vocabulary_masks()
+        # Move masks to same device as model
+        device = next(self.parameters()).device
+        self.vocab_masks = {k: v.to(device) for k, v in self.vocab_masks.items()}
+        self.use_constrained_decoding = True
+        print(f"Structured generation enabled with {len(self.vocab_masks)} mask types")
+    
+    def _get_allowed_tokens_mask(self, input_ids: torch.Tensor) -> torch.Tensor:
+        """Get a mask of allowed tokens based on current generation state.
+        
+        Args:
+            input_ids: Tensor of shape (batch_size, seq_len).
+            
+        Returns:
+            Boolean tensor of shape (batch_size, vocab_size) where True = allowed.
+        """
+        # Auto-prepare if not already done (for compatibility with evaluate.py)
+        if self.vocab_masks is None and self.tokenizer_ref is None:
+            # Try to auto-prepare using component mode tokenizer
+            try:
+                from tokenizer import ChessTokenizer
+                tokenizer = ChessTokenizer.build_vocab_more_detailed()
+                self.prepare_for_structured_generation(tokenizer)
+            except:
+                # If auto-prepare fails, allow all tokens (no constraints)
+                device = input_ids.device
+                return torch.ones((input_ids.shape[0], self.config.vocab_size), dtype=torch.bool, device=device)
+        
+        if self.vocab_masks is None or self.tokenizer_ref is None:
+            # Still not prepared, allow all tokens
+            device = input_ids.device
+            return torch.ones((input_ids.shape[0], self.config.vocab_size), dtype=torch.bool, device=device)
+        
+        batch_size = input_ids.shape[0]
+        vocab_size = self.config.vocab_size
+        device = input_ids.device
+        
+        # Analyze generation state
+        state = self.tokenizer_ref.analyze_generation_state(input_ids)
+        
+        # Handle single batch (state is dict) vs multi-batch (state is list)
+        if isinstance(state, dict):
+            states = [state]
+        else:
+            states = state
+        
+        # Build mask for each batch element
+        masks = []
+        for s in states:
+            position = s['position']
+            expected_color = s['expected_color']
+            
+            # Initialize mask (all False)
+            mask = torch.zeros(vocab_size, dtype=torch.bool, device=device)
+            
+            if position == 0:
+                # Expect piece token with correct color
+                if expected_color == 'W':
+                    mask = self.vocab_masks['white_piece'].clone()
+                else:
+                    mask = self.vocab_masks['black_piece'].clone()
+            elif position == 1:
+                # Expect from_square
+                mask = self.vocab_masks['square'].clone()
+            elif position == 2:
+                # Expect to_square
+                mask = self.vocab_masks['square'].clone()
+            else:  # position == 3
+                # Expect suffix or [EOM]
+                mask = self.vocab_masks['suffix'] | self.vocab_masks['eom']
+            
+            masks.append(mask)
+        
+        # Stack into (batch_size, vocab_size)
+        return torch.stack(masks, dim=0)
+    
+    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)
+        
+        # Apply structured generation mask during inference (eval mode without labels)
+        if not self.training and labels is None and hasattr(self, 'use_constrained_decoding') and self.use_constrained_decoding:
+            # Get mask for valid next tokens based on generation state
+            allowed_mask = self._get_allowed_tokens_mask(input_ids)  # (batch_size, vocab_size)
+            # Apply mask to last position logits only (where we're generating)
+            last_logits = logits[:, -1, :]  # (batch_size, vocab_size)
+            # Set disallowed tokens to -inf
+            last_logits = last_logits.masked_fill(~allowed_mask, float('-inf'))
+            # Update logits
+            logits[:, -1, :] = last_logits
+        
+        # Compute loss if labels are provided
+        loss = None
+        if labels is not None:
+            # Shift logits and labels for next-token prediction
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            
+            # Flatten for cross-entropy
+            loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
+            # loss_fct = nn.CrossEntropyLoss(ignore_index=self.config.pad_token_id)
+            loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)),
+                shift_labels.view(-1),
+            )
+        
+        if not return_dict:
+            output = (logits,)
+            return ((loss,) + output) if loss is not None else output
+        
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=None,
+            hidden_states=None,
+            attentions=None,
+        )
+    
+    @torch.no_grad()
+    def generate_move(
+        self,
+        input_ids: torch.LongTensor,
+        temperature: float = 1.0,
+        top_k: Optional[int] = None,
+        top_p: Optional[float] = None,
+    ) -> int:
+        """
+        Generate the next move given a sequence of moves.
+        
+        Args:
+            input_ids: Token IDs of shape (1, seq_len).
+            temperature: Sampling temperature (1.0 = no change).
+            top_k: If set, only sample from top k tokens.
+            top_p: If set, use nucleus sampling with this threshold.
+        
+        Returns:
+            The token ID of the predicted next move.
+        """
+        self.eval()
+        
+        # Get logits for the last position
+        outputs = self(input_ids)
+        logits = outputs.logits[:, -1, :] / temperature
+        
+        # Apply top-k filtering
+        if top_k is not None:
+            indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+            logits[indices_to_remove] = float("-inf")
+        
+        # Apply top-p (nucleus) filtering
+        if top_p is not None:
+            sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+            cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+            
+            # Remove tokens with cumulative probability above the threshold
+            sorted_indices_to_remove = cumulative_probs > top_p
+            sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+            sorted_indices_to_remove[..., 0] = 0
+            
+            indices_to_remove = sorted_indices_to_remove.scatter(
+                dim=-1, index=sorted_indices, src=sorted_indices_to_remove
+            )
+            logits[indices_to_remove] = float("-inf")
+        
+        # Sample from the distribution
+        probs = F.softmax(logits, dim=-1)
+        next_token = torch.multinomial(probs, num_samples=1)
+        
+        return next_token.item()
+
+
+# Register the model with Auto classes for easy loading
+from transformers import AutoConfig, AutoModelForCausalLM
+
+AutoConfig.register("chess_transformer", ChessConfig)
+AutoModelForCausalLM.register(ChessConfig, ChessForCausalLM)

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72437bd478fb1d15478bb2d4079a277d8fdec02823433d900c164e77e5bc19de
+size 3492656

special_tokens_map.json

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

tokenizer.py

@@ -0,0 +1,571 @@
+"""
+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]"
+    EOM_TOKEN = "[EOM]"  # End of Move - marks boundary between moves
+    
+    def __init__(
+        self,
+        vocab_file: Optional[str] = None,
+        vocab: Optional[Dict[str, int]] = None,
+        component_mode: bool = False,
+        **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).
+            component_mode: If True, tokenize moves into components (WP, e2, e4).
+            **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
+        self._eom_token = self.EOM_TOKEN
+        
+        # Component mode flag (for splitting moves into parts)
+        self._component_mode = component_mode
+
+        # 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)
+        kwargs.pop("eom_token", None)
+        kwargs.pop("component_mode", 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,
+            component_mode=component_mode,  # This gets saved to tokenizer_config.json
+            **kwargs,
+        )
+        # Store EOM token ID for easy access
+        self.eom_token_id = self._vocab.get(self.EOM_TOKEN, -1)
+    
+    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)
+    
+    @classmethod
+    def build_vocab_more_detailed(
+        cls,
+        ) -> "ChessTokenizer":
+        """
+        Build a component-based tokenizer for chess moves.
+        
+        Instead of one token per move (WPe2e4), splits into components:
+        WPe2e4 -> [WP, e2, e4]
+        BNg8f6(x) -> [BN, g8, f6, (x)]
+        
+        This gives ~90 tokens instead of ~1200, with better generalization.
+        
+        Returns:
+            A ChessTokenizer with component vocabulary.
+        """
+        # Combined color+piece tokens (avoids B collision between Black and Bishop)
+        tokens_pieces = [
+            "WP", "WN", "WB", "WR", "WQ", "WK",  # White pieces
+            "BP", "BN", "BB", "BR", "BQ", "BK",  # Black pieces
+        ]
+        
+        # the positions:
+        files = ['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h']
+        ranks = ['1', '2', '3', '4', '5', '6', '7', '8']
+        tokens_positions = [f + r for f in files for r in ranks]
+        
+        # the special suffixes:
+        tokens_suffixes = [
+            "(x)",   # capture
+            "(+)",   # check
+            "(x+)",  # capture + check
+            "(+*)",  # checkmate
+            "(x+*)", # capture + checkmate
+            "(o)",   # kingside castling
+            "(O)",   # queenside castling
+            "(xE)",  # en passant
+            "=Q",    # promotion to queen
+            "=R",    # promotion to rook
+            "=B",    # promotion to bishop
+            "=N",    # promotion to knight
+        ]
+        
+        # Combine all tokens
+        tokens = tokens_pieces + tokens_positions + tokens_suffixes
+        
+        # Build vocabulary with [EOM] for move boundaries
+        # [EOM] helps the model understand when a move ends
+        special_tokens = [cls.PAD_TOKEN, cls.BOS_TOKEN, cls.EOS_TOKEN, cls.UNK_TOKEN, cls.EOM_TOKEN]
+        vocab = {token: idx for idx, token in enumerate(special_tokens + tokens)}
+        for ind, token in enumerate(special_tokens+tokens):
+            print(f"Token {ind}: {token}")
+        # Pass component_mode=True so it gets saved to tokenizer_config.json
+        return cls(vocab=vocab, component_mode=True)
+  
+    @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.
+        
+        If component_mode is enabled, splits each move into parts:
+        WPe2e4 -> [W, P, e2, e4, " "]
+        BNg8f6(x) -> [B, N, g8, f6, (x), " "]
+        
+        Args:
+            text: A string of space-separated moves.
+        
+        Returns:
+            List of tokens.
+        """
+        if getattr(self, '_component_mode', False):
+            return self._tokenize_components(text)
+        return text.strip().split()
+    
+    def _tokenize_components(self, text: str) -> List[str]:
+        """
+        Tokenize moves into component parts with [EOM] boundaries.
+        
+        Move format: [Color][Piece][from_square][to_square][suffix] [EOM]
+        Example: 
+          WPe2e4 -> [WP, e2, e4, EOM]
+          BNg8f6(x) -> [BN, g8, f6, (x), EOM]
+        """
+        import re
+        
+        tokens = []
+        moves = text.strip().split()
+        
+        for i, move in enumerate(moves):
+            # Skip special tokens
+            if move in [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN, self.EOM_TOKEN]:
+                tokens.append(move)
+                continue
+            
+            # Parse move: ColorPiece + from_square + to_square + optional suffix
+            # Pattern: (W|B)(P|N|B|R|Q|K)([a-h][1-8])([a-h][1-8])(suffix)?
+            pattern = r'^([WB])([PNBRQK])([a-h][1-8])([a-h][1-8])(.*)$'
+            match = re.match(pattern, move)
+            
+            if match:
+                color, piece, from_sq, to_sq, suffix = match.groups()
+                # Combined color+piece token (e.g., "WP", "BN", "BB")
+                tokens.append(color + piece)
+                tokens.extend([from_sq, to_sq])
+                
+                # Handle suffix (could be combination like "(x+)" or "=Q")
+                if suffix:
+                    # Try to match known suffixes
+                    suffix_pattern = r'(\(x\+\*\)|\(x\+\)|\(\+\*\)|\(xE\)|\(x\)|\(\+\)|\(o\)|\(O\)|=Q|=R|=B|=N)'
+                    suffix_matches = re.findall(suffix_pattern, suffix)
+                    tokens.extend(suffix_matches)
+                
+                # Add [EOM] to mark end of this move
+                tokens.append(self.EOM_TOKEN)
+            else:
+                # Fallback: add as unknown + EOM
+                tokens.append(self.UNK_TOKEN)
+                tokens.append(self.EOM_TOKEN)
+        
+        return tokens
+    
+    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."""
+        token = self._ids_to_tokens.get(index, self.UNK_TOKEN)
+        # Convert [EOM] to whitespace for evaluator compatibility
+        # This makes _generate_until_whitespace stop after one move
+        if token == self.EOM_TOKEN:
+            return " "
+        return token
+    
+    # Color+piece tokens that mark the start of a new move
+    _MOVE_START_TOKENS = {"WP", "WN", "WB", "WR", "WQ", "WK", "BP", "BN", "BB", "BR", "BQ", "BK"}
+    
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        """Convert a list of tokens back to a string.
+        
+        In component mode, reconstructs moves by replacing [EOM] with spaces.
+        CRITICAL: [EOM] must decode to a non-empty whitespace string so that
+        the evaluator's _generate_until_whitespace stops after one move.
+        """
+        # Filter out special tokens except EOM for cleaner output
+        special = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}
+        
+        if getattr(self, '_component_mode', False):
+            # Reconstruct with [EOM] as space delimiter
+            result = []
+            for token in tokens:
+                if token == self.EOM_TOKEN:
+                    # MUST be non-empty whitespace for evaluator
+                    result.append(" ")
+                elif token not in special:
+                    result.append(token)
+            # Don't strip! We need the trailing space from [EOM]
+            return "".join(result)
+        
+        # Non-component mode: just join with spaces
+        filtered = [t for t in tokens if t not in special]
+        return " ".join(filtered)
+    
+    # =========================================================================
+    # Structured Generation Support Methods
+    # =========================================================================
+    
+    def get_token_category(self, token: str) -> str:
+        """Categorize a token into: piece, square, suffix, eom, or special.
+        
+        Args:
+            token: Token string to categorize.
+            
+        Returns:
+            Category name: 'piece', 'square', 'suffix', 'eom', or 'special'.
+        """
+        if token in [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]:
+            return 'special'
+        if token == self.EOM_TOKEN:
+            return 'eom'
+        if self.is_piece_token(token):
+            return 'piece'
+        if self.is_square_token(token):
+            return 'square'
+        if self.is_suffix_token(token):
+            return 'suffix'
+        return 'unknown'
+    
+    def is_piece_token(self, token: str) -> bool:
+        """Check if token is a piece token (WP, BN, etc.)."""
+        return token in ['WP', 'WN', 'WB', 'WR', 'WQ', 'WK', 'BP', 'BN', 'BB', 'BR', 'BQ', 'BK']
+    
+    def is_square_token(self, token: str) -> bool:
+        """Check if token is a square token (e2, g8, etc.)."""
+        if len(token) != 2:
+            return False
+        return token[0] in 'abcdefgh' and token[1] in '12345678'
+    
+    def is_suffix_token(self, token: str) -> bool:
+        """Check if token is a suffix token ((x), (+), =Q, etc.)."""
+        return token in ['(x)', '(+)', '(x+)', '(+*)', '(x+*)', '(o)', '(O)', '(xE)', '=Q', '=R', '=B', '=N']
+    
+    def is_eom_token(self, token: str) -> bool:
+        """Check if token is the [EOM] token."""
+        return token == self.EOM_TOKEN
+    
+    def get_token_color(self, token: str) -> Optional[str]:
+        """Get the color ('W' or 'B') from a piece token, None otherwise."""
+        if self.is_piece_token(token) and len(token) >= 2:
+            return token[0]  # 'W' or 'B'
+        return None
+    
+    def build_vocabulary_masks(self) -> dict:
+        """Build boolean masks for each token category.
+        
+        Returns:
+            Dictionary with keys: 'piece', 'square', 'suffix', 'eom', 'white_piece', 'black_piece'.
+            Each value is a boolean list/tensor of length vocab_size.
+        """
+        import torch
+        
+        vocab_size = len(self._vocab)
+        masks = {
+            'piece': [False] * vocab_size,
+            'square': [False] * vocab_size,
+            'suffix': [False] * vocab_size,
+            'eom': [False] * vocab_size,
+            'white_piece': [False] * vocab_size,
+            'black_piece': [False] * vocab_size,
+        }
+        
+        for token, token_id in self._vocab.items():
+            if self.is_piece_token(token):
+                masks['piece'][token_id] = True
+                color = self.get_token_color(token)
+                if color == 'W':
+                    masks['white_piece'][token_id] = True
+                elif color == 'B':
+                    masks['black_piece'][token_id] = True
+            elif self.is_square_token(token):
+                masks['square'][token_id] = True
+            elif self.is_suffix_token(token):
+                masks['suffix'][token_id] = True
+            elif self.is_eom_token(token):
+                masks['eom'][token_id] = True
+        
+        # Convert to tensors
+        return {k: torch.tensor(v, dtype=torch.bool) for k, v in masks.items()}
+    
+    def analyze_generation_state(self, input_ids: torch.Tensor) -> dict:
+        """Analyze the current generation state to determine next expected token.
+        
+        Args:
+            input_ids: Tensor of shape (batch_size, seq_len) with token IDs.
+            
+        Returns:
+            Dictionary with:
+            - 'position': 0 (piece), 1 (from_square), 2 (to_square), 3 (suffix/eom)
+            - 'expected_color': 'W' or 'B'
+            - 'last_eom_idx': Index of last [EOM] token in sequence
+        """
+        batch_size = input_ids.shape[0]
+        results = []
+        
+        for b in range(batch_size):
+            seq = input_ids[b].tolist()
+            
+            # Find last [EOM] or [BOS]
+            last_eom_idx = -1
+            for i in range(len(seq) - 1, -1, -1):
+                token = self._ids_to_tokens.get(seq[i], self.UNK_TOKEN)
+                if token in [self.EOM_TOKEN, self.BOS_TOKEN]:
+                    last_eom_idx = i
+                    break
+            
+            # Count tokens since last [EOM]/[BOS] (excluding padding)
+            tokens_since_boundary = []
+            for i in range(last_eom_idx + 1, len(seq)):
+                token = self._ids_to_tokens.get(seq[i], self.UNK_TOKEN)
+                if token != self.PAD_TOKEN:
+                    tokens_since_boundary.append(token)
+            
+            # Determine position in move structure: [Piece][Square][Square][Suffix?][EOM]
+            num_tokens = len(tokens_since_boundary)
+            
+            if num_tokens == 0:
+                position = 0  # Expect piece
+            elif num_tokens == 1:
+                position = 1  # Expect from_square
+            elif num_tokens == 2:
+                position = 2  # Expect to_square
+            else:
+                position = 3  # Expect suffix or [EOM]
+            
+            # Determine expected color by counting complete moves
+            # Count [EOM] tokens to get move number
+            eom_count = sum(1 for i in seq if self._ids_to_tokens.get(i, '') == self.EOM_TOKEN)
+            expected_color = 'W' if eom_count % 2 == 0 else 'B'
+            
+            results.append({
+                'position': position,
+                'expected_color': expected_color,
+                'last_eom_idx': last_eom_idx,
+            })
+        
+        # For single batch, return dict directly; for multi-batch, return list
+        return results[0] if batch_size == 1 else results
+    
+    def save_vocabulary(
+        self,
+        save_directory: str,
+        filename_prefix: Optional[str] = None,
+    ) -> tuple:
+        """
+        Save the vocabulary to a JSON file.
+        
+        Args:
+            save_directory: Directory to save the vocabulary.
+            filename_prefix: Optional prefix for the filename.
+        
+        Returns:
+            Tuple containing the path to the saved vocabulary file.
+        """
+        if not os.path.isdir(save_directory):
+            os.makedirs(save_directory, exist_ok=True)
+        
+        vocab_file = os.path.join(
+            save_directory,
+            (filename_prefix + "-" if filename_prefix else "") + "vocab.json",
+        )
+        
+        with open(vocab_file, "w", encoding="utf-8") as f:
+            json.dump(self._vocab, f, ensure_ascii=False, indent=2)
+        
+        return (vocab_file,)
+
+
+def count_vocab_from_dataset(
+    dataset_name: str = "dlouapre/lichess_2025-01_1M",
+    split: str = "train",
+    column: str = "text",
+    max_samples: Optional[int] = 10000,
+) -> Dict[str, int]:
+    """
+    Count token frequencies in a dataset (useful for vocabulary analysis).
+    
+    Args:
+        dataset_name: Name of the dataset on Hugging Face Hub.
+        split: Dataset split to use.
+        column: Column containing the game strings.
+        max_samples: Maximum number of samples to process.
+    
+    Returns:
+        Dictionary mapping tokens to their frequencies.
+    """
+    from collections import Counter
+    from datasets import load_dataset
+    
+    dataset = load_dataset(dataset_name, split=split)
+    
+    if max_samples is not None:
+        dataset = dataset.select(range(min(max_samples, len(dataset))))
+    
+    token_counts = Counter()
+    
+    for example in dataset:
+        moves = example[column].strip().split()
+        token_counts.update(moves)
+    
+    return dict(token_counts)

tokenizer_config.json

@@ -0,0 +1,51 @@
+{
+  "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
+    }
+  },
+  "bos_token": "[BOS]",
+  "clean_up_tokenization_spaces": false,
+  "component_mode": true,
+  "eos_token": "[EOS]",
+  "extra_special_tokens": {},
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "[PAD]",
+  "tokenizer_class": "ChessTokenizer",
+  "unk_token": "[UNK]",
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenizer.ChessTokenizer",
+      null
+    ]
+  }
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:078339b1e2d26d1eafa619e3e9eaad471f271142e053207e14cd496ff6660728
+size 5713

vocab.json

@@ -0,0 +1,95 @@
+{
+  "[PAD]": 0,
+  "[BOS]": 1,
+  "[EOS]": 2,
+  "[UNK]": 3,
+  "[EOM]": 4,
+  "WP": 5,
+  "WN": 6,
+  "WB": 7,
+  "WR": 8,
+  "WQ": 9,
+  "WK": 10,
+  "BP": 11,
+  "BN": 12,
+  "BB": 13,
+  "BR": 14,
+  "BQ": 15,
+  "BK": 16,
+  "a1": 17,
+  "a2": 18,
+  "a3": 19,
+  "a4": 20,
+  "a5": 21,
+  "a6": 22,
+  "a7": 23,
+  "a8": 24,
+  "b1": 25,
+  "b2": 26,
+  "b3": 27,
+  "b4": 28,
+  "b5": 29,
+  "b6": 30,
+  "b7": 31,
+  "b8": 32,
+  "c1": 33,
+  "c2": 34,
+  "c3": 35,
+  "c4": 36,
+  "c5": 37,
+  "c6": 38,
+  "c7": 39,
+  "c8": 40,
+  "d1": 41,
+  "d2": 42,
+  "d3": 43,
+  "d4": 44,
+  "d5": 45,
+  "d6": 46,
+  "d7": 47,
+  "d8": 48,
+  "e1": 49,
+  "e2": 50,
+  "e3": 51,
+  "e4": 52,
+  "e5": 53,
+  "e6": 54,
+  "e7": 55,
+  "e8": 56,
+  "f1": 57,
+  "f2": 58,
+  "f3": 59,
+  "f4": 60,
+  "f5": 61,
+  "f6": 62,
+  "f7": 63,
+  "f8": 64,
+  "g1": 65,
+  "g2": 66,
+  "g3": 67,
+  "g4": 68,
+  "g5": 69,
+  "g6": 70,
+  "g7": 71,
+  "g8": 72,
+  "h1": 73,
+  "h2": 74,
+  "h3": 75,
+  "h4": 76,
+  "h5": 77,
+  "h6": 78,
+  "h7": 79,
+  "h8": 80,
+  "(x)": 81,
+  "(+)": 82,
+  "(x+)": 83,
+  "(+*)": 84,
+  "(x+*)": 85,
+  "(o)": 86,
+  "(O)": 87,
+  "(xE)": 88,
+  "=Q": 89,
+  "=R": 90,
+  "=B": 91,
+  "=N": 92
+}