Chess Challenge submission by SmileChou

README.md

@@ -0,0 +1,31 @@
+---
+library_name: transformers
+tags:
+- chess
+- llm-course
+- chess-challenge
+license: mit
+---
+
+# smile_chess_chanllenge_done
+
+Chess model submitted to the LLM Course Chess Challenge.
+
+## Submission Info
+
+- **Submitted by**: [SmileChou](https://huggingface.co/SmileChou)
+- **Parameters**: 997,964
+- **Organization**: LLM-course
+
+## Usage
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model = AutoModelForCausalLM.from_pretrained("LLM-course/smile_chess_chanllenge_done", trust_remote_code=True)
+tokenizer = AutoTokenizer.from_pretrained("LLM-course/smile_chess_chanllenge_done", 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 +1,25 @@
+{
+  "architectures": [
+    "ChessForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "model.ChessConfig",
+    "AutoModelForCausalLM": "model.ChessForCausalLM"
+  },
+
+  "bos_token_id": 1,
+  "dropout": 0.05,
+  "dtype": "float32",
+  "eos_token_id": 2,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "chess_transformer",
+  "n_ctx": 256,
+  "n_embd": 128,
+  "n_head": 8,
+  "n_inner": 354,
+  "n_layer": 6,
+  "pad_token_id": 0,
+  "tie_weights": true,
+  "transformers_version": "4.57.6",
+  "vocab_size": 148
+}

model.py

@@ -0,0 +1,437 @@
+"""
+Chess Transformer Model for the Chess Challenge.
+
+This module provides a simple GPT-style transformer architecture
+designed to fit within the 1M parameter constraint.
+
+Key components:
+- ChessConfig: Configuration class for model hyperparameters
+- ChessForCausalLM: The main model class for next-move prediction
+"""
+
+from __future__ import annotations
+
+import math
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import PretrainedConfig, PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+
+class ChessConfig(PretrainedConfig):
+    """
+    Configuration class for the Chess Transformer model.
+    
+    This configuration is designed for a ~1M parameter model.
+    Students can adjust these values to explore different architectures.
+    
+    Parameter budget breakdown (with default values):
+    - Embeddings (vocab): 1200 x 128 = 153,600
+    - Position Embeddings: 256 x 128 = 32,768
+    - Transformer Layers: 6 x ~120,000 = ~720,000
+    - LM Head (with weight tying): 0 (shared with embeddings)
+    - Total: ~906,000 parameters
+    
+    Attributes:
+        vocab_size: Size of the vocabulary (number of unique moves).
+        n_embd: Embedding dimension (d_model).
+        n_layer: Number of transformer layers.
+        n_head: Number of attention heads.
+        n_ctx: Maximum sequence length (context window).
+        n_inner: Feed-forward inner dimension (default: 3 * n_embd).
+        dropout: Dropout probability.
+        layer_norm_epsilon: Epsilon for layer normalization.
+        tie_weights: Whether to tie embedding and output weights.
+    """
+    
+    model_type = "chess_transformer"
+    
+    def __init__(
+        self,
+        vocab_size: int = 1200,
+        n_embd: int = 128,
+        n_layer: int = 6,
+        n_head: int = 4,
+        n_ctx: int = 256,
+        n_inner: Optional[int] = None,
+        dropout: float = 0.05,
+        layer_norm_epsilon: float = 1e-5,
+        tie_weights: bool = True,
+        pad_token_id: int = 0,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        **kwargs,
+    ):
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            **kwargs,
+        )
+        
+        self.vocab_size = vocab_size
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.n_ctx = n_ctx
+        self.n_inner = n_inner if n_inner is not None else 3 * n_embd  # Reduced from 4x to 3x
+        self.dropout = dropout
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.tie_weights = tie_weights
+        # Inform HF base class about tying behavior
+        self.tie_word_embeddings = bool(tie_weights)
+
+
+class MultiHeadAttention(nn.Module):
+    """
+    Multi-head self-attention module.
+    
+    This is a standard scaled dot-product attention implementation
+    with causal masking for autoregressive generation.
+    """
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        
+        assert config.n_embd % config.n_head == 0, \
+            f"n_embd ({config.n_embd}) must be divisible by n_head ({config.n_head})"
+        
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.head_dim = config.n_embd // config.n_head
+        
+        # Combined QKV projection for efficiency
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd)
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd)
+        
+        self.dropout = nn.Dropout(config.dropout)
+        
+        # Causal mask (will be created on first forward pass)
+        self.register_buffer(
+            "bias",
+            torch.tril(torch.ones(config.n_ctx, config.n_ctx)).view(
+                1, 1, config.n_ctx, config.n_ctx
+            ),
+            persistent=False,
+        )
+    
+    def forward(
+        self,
+        x: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, seq_len, _ = x.size()
+        
+        # Compute Q, K, V
+        qkv = self.c_attn(x)
+        q, k, v = qkv.split(self.n_embd, dim=2)
+        
+        # Reshape for multi-head attention
+        q = q.view(batch_size, seq_len, self.n_head, self.head_dim).transpose(1, 2)
+        k = k.view(batch_size, seq_len, self.n_head, self.head_dim).transpose(1, 2)
+        v = v.view(batch_size, seq_len, self.n_head, self.head_dim).transpose(1, 2)
+        
+        # Scaled dot-product attention
+        attn_weights = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
+        
+        # Apply causal mask
+        causal_mask = self.bias[:, :, :seq_len, :seq_len]
+        attn_weights = attn_weights.masked_fill(causal_mask == 0, float("-inf"))
+        
+        # Apply attention mask (for padding)
+        if attention_mask is not None:
+            # attention_mask shape: (batch_size, seq_len) -> (batch_size, 1, 1, seq_len)
+            attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+            attn_weights = attn_weights.masked_fill(attention_mask == 0, float("-inf"))
+        
+        attn_weights = F.softmax(attn_weights, dim=-1)
+        attn_weights = self.dropout(attn_weights)
+        
+        # Apply attention to values
+        attn_output = torch.matmul(attn_weights, v)
+        
+        # Reshape back
+        attn_output = attn_output.transpose(1, 2).contiguous().view(
+            batch_size, seq_len, self.n_embd
+        )
+        
+        # Output projection
+        attn_output = self.c_proj(attn_output)
+        
+        return attn_output
+
+
+class FeedForward(nn.Module):
+    """
+    Feed-forward network (MLP) module.
+    
+    Standard two-layer MLP with GELU activation.
+    """
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        
+        self.c_fc = nn.Linear(config.n_embd, config.n_inner)
+        self.c_proj = nn.Linear(config.n_inner, config.n_embd)
+        self.dropout = nn.Dropout(config.dropout)
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.c_fc(x)
+        x = F.gelu(x)
+        x = self.c_proj(x)
+        x = self.dropout(x)
+        return x
+
+
+class TransformerBlock(nn.Module):
+    """
+    A single transformer block with attention and feed-forward layers.
+    
+    Uses pre-normalization (LayerNorm before attention/FFN) for better
+    training stability.
+    """
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        
+        self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.attn = MultiHeadAttention(config)
+        self.ln_2 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.mlp = FeedForward(config)
+    
+    def forward(
+        self,
+        x: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        # Pre-norm attention
+        x = x + self.attn(self.ln_1(x), attention_mask=attention_mask)
+        # Pre-norm FFN
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class ChessForCausalLM(PreTrainedModel):
+    """
+    Chess Transformer for Causal Language Modeling (next-move prediction).
+    
+    This model is designed to predict the next chess move given a sequence
+    of previous moves. It uses a GPT-style architecture with:
+    - Token embeddings for chess moves
+    - Learned positional embeddings
+    - Stacked transformer blocks
+    - Linear head for next-token prediction
+    
+    The model supports weight tying between the embedding layer and the
+    output projection to save parameters.
+    
+    Example:
+        >>> config = ChessConfig(vocab_size=1200, n_embd=128, n_layer=6)
+        >>> model = ChessForCausalLM(config)
+        >>> inputs = {"input_ids": torch.tensor([[1, 42, 87]])}
+        >>> outputs = model(**inputs)
+        >>> next_move_logits = outputs.logits[:, -1, :]
+    """
+    
+    config_class = ChessConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    # Suppress missing-key warning for tied lm_head when loading
+    keys_to_ignore_on_load_missing = ["lm_head.weight"]
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__(config)
+        
+        # Token and position embeddings
+        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
+        self.wpe = nn.Embedding(config.n_ctx, config.n_embd)
+        
+        self.drop = nn.Dropout(config.dropout)
+        
+        # Transformer blocks
+        self.h = nn.ModuleList([
+            TransformerBlock(config) for _ in range(config.n_layer)
+        ])
+        
+        # Final layer norm
+        self.ln_f = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        
+        # Output head
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+        
+        # Declare tied weights for proper serialization
+        if config.tie_weights:
+            self._tied_weights_keys = ["lm_head.weight"]
+        
+        # Initialize weights
+        self.post_init()
+        
+        # Tie weights if configured
+        if config.tie_weights:
+            self.tie_weights()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.wte
+
+    def set_input_embeddings(self, new_embeddings: nn.Module):
+        self.wte = new_embeddings
+        if getattr(self.config, "tie_weights", False):
+            self.tie_weights()
+
+    def get_output_embeddings(self) -> nn.Module:
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings: nn.Module):
+        self.lm_head = new_embeddings
+
+    def tie_weights(self):
+        # Use HF helper to tie or clone depending on config
+        if getattr(self.config, "tie_weights", False) or getattr(self.config, "tie_word_embeddings", False):
+            self._tie_or_clone_weights(self.lm_head, self.wte)
+    
+    def _init_weights(self, module: nn.Module):
+        """Initialize weights following GPT-2 style."""
+        if isinstance(module, nn.Linear):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
+            if module.bias is not None:
+                torch.nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
+        elif isinstance(module, nn.LayerNorm):
+            torch.nn.init.ones_(module.weight)
+            torch.nn.init.zeros_(module.bias)
+    
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        """
+        Forward pass of the model.
+        
+        Args:
+            input_ids: Token IDs of shape (batch_size, seq_len).
+            attention_mask: Attention mask of shape (batch_size, seq_len).
+            position_ids: Position IDs of shape (batch_size, seq_len).
+            labels: Labels for language modeling loss.
+            return_dict: Whether to return a ModelOutput object.
+        
+        Returns:
+            CausalLMOutputWithPast containing loss (if labels provided) and logits.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        
+        batch_size, seq_len = input_ids.size()
+        device = input_ids.device
+        
+        # Create position IDs if not provided
+        if position_ids is None:
+            position_ids = torch.arange(seq_len, device=device).unsqueeze(0).expand(batch_size, -1)
+        
+        # Get embeddings
+        token_embeds = self.wte(input_ids)
+        position_embeds = self.wpe(position_ids)
+        hidden_states = self.drop(token_embeds + position_embeds)
+        
+        # Pass through transformer blocks
+        for block in self.h:
+            hidden_states = block(hidden_states, attention_mask=attention_mask)
+        
+        # Final layer norm
+        hidden_states = self.ln_f(hidden_states)
+        
+        # Get logits
+        logits = self.lm_head(hidden_states)
+        
+        # Compute loss if labels are provided
+        loss = None
+        if labels is not None:
+            # Shift logits and labels for next-token prediction
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            
+            # Flatten for cross-entropy
+            loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
+            loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)),
+                shift_labels.view(-1),
+            )
+        
+        if not return_dict:
+            output = (logits,)
+            return ((loss,) + output) if loss is not None else output
+        
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=None,
+            hidden_states=None,
+            attentions=None,
+        )
+    
+    @torch.no_grad()
+    def generate_move(
+        self,
+        input_ids: torch.LongTensor,
+        temperature: float = 1.0,
+        top_k: Optional[int] = None,
+        top_p: Optional[float] = None,
+    ) -> int:
+        """
+        Generate the next move given a sequence of moves.
+        
+        Args:
+            input_ids: Token IDs of shape (1, seq_len).
+            temperature: Sampling temperature (1.0 = no change).
+            top_k: If set, only sample from top k tokens.
+            top_p: If set, use nucleus sampling with this threshold.
+        
+        Returns:
+            The token ID of the predicted next move.
+        """
+        self.eval()
+        
+        # Get logits for the last position
+        outputs = self(input_ids)
+        logits = outputs.logits[:, -1, :] / temperature
+        
+        # Apply top-k filtering
+        if top_k is not None:
+            indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+            logits[indices_to_remove] = float("-inf")
+        
+        # Apply top-p (nucleus) filtering
+        if top_p is not None:
+            sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+            cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+            
+            # Remove tokens with cumulative probability above the threshold
+            sorted_indices_to_remove = cumulative_probs > top_p
+            sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+            sorted_indices_to_remove[..., 0] = 0
+            
+            indices_to_remove = sorted_indices_to_remove.scatter(
+                dim=-1, index=sorted_indices, src=sorted_indices_to_remove
+            )
+            logits[indices_to_remove] = float("-inf")
+        
+        # Sample from the distribution
+        probs = F.softmax(logits, dim=-1)
+        next_token = torch.multinomial(probs, num_samples=1)
+        
+        return next_token.item()
+
+
+# Register the model with Auto classes for easy loading
+from transformers import AutoConfig, AutoModelForCausalLM
+
+AutoConfig.register("chess_transformer", ChessConfig)
+AutoModelForCausalLM.register(ChessConfig, ChessForCausalLM)

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1b50a2c130d27f4770a8fc5c2ffe444f4610ad219da6d53e2e362319be33c57
+size 3998304

special_tokens_map.json

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

tokenizer.py

@@ -0,0 +1,182 @@
+"""
+ Chess Tokenizer (Refactored).
+
+Architecture:
+- Splits chess moves into atomic component tokens.
+- Structure: [Actor] -> [Source_Square] -> [Target_Square] -> [Promotion?]
+- Output format: "WP", "e2_f", "e4_t"
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+from typing import Dict, List, Optional, Any, Tuple
+
+from transformers import PreTrainedTokenizer
+
+
+class ChessTokenizer(PreTrainedTokenizer):
+    """
+    A tokenizer that breaks chess moves into explicit actor and coordinate tokens.
+    Designed for high-precision state tracking.
+    """
+    
+    model_input_names = ["input_ids", "attention_mask"]
+    vocab_files_names = {"vocab_file": "vocab.json"}
+
+    # --- Configuration Constants ---
+    TOKENS_SPECIAL = ["[PAD]", "[BOS]", "[EOS]", "[UNK]"]
+    CHARS_PIECE = "PNBRQK"
+    CHARS_COLOR = "WB"
+    CHARS_FILE = "abcdefgh"
+    CHARS_RANK = "12345678"
+    CHARS_PROMO = {"q", "r", "b", "n"}
+
+    # Regex to validate and parse standard Lichess moves (e.g., WPe2e4)
+    # Group 1: Color, 2: Piece, 3: Source, 4: Target, 5: Suffix
+    PATTERN_MOVE = 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: Any,
+    ):
+        # Initialize special tokens for the parent class
+        self._pad_token = self.TOKENS_SPECIAL[0]
+        self._bos_token = self.TOKENS_SPECIAL[1]
+        self._eos_token = self.TOKENS_SPECIAL[2]
+        self._unk_token = self.TOKENS_SPECIAL[3]
+
+        # Clean kwargs to prevent collisions
+        for token_arg in ["pad_token", "bos_token", "eos_token", "unk_token"]:
+            kwargs.pop(token_arg, None)
+
+        # 1. Load Vocabulary
+        if vocab:
+            self._vocab = vocab
+        elif vocab_file and os.path.isfile(vocab_file):
+            with open(vocab_file, "r", encoding="utf-8") as f:
+                self._vocab = json.load(f)
+        else:
+            self._vocab = self._generate_vocabulary()
+
+        # 2. Build ID-to-Token Map
+        self._id_to_token = {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,
+        )
+
+    def _generate_vocabulary(self) -> Dict[str, int]:
+        """Constructs the fixed dictionary of tokens."""
+        token_list = list(self.TOKENS_SPECIAL)
+
+        # A. Actor Tokens (e.g., WP, BN)
+        token_list.extend(
+            f"{c}{p}" for c in self.CHARS_COLOR for p in self.CHARS_PIECE
+        )
+
+        # B. Coordinate Tokens (Source & Target)
+        squares = [f"{f}{r}" for r in self.CHARS_RANK for f in self.CHARS_FILE]
+        token_list.extend(f"{sq}_f" for sq in squares) # From
+        token_list.extend(f"{sq}_t" for sq in squares) # To
+
+        # C. Promotion Tokens (Sorted for consistency)
+        token_list.extend(sorted(self.CHARS_PROMO))
+
+        return {token: idx for idx, token in enumerate(token_list)}
+
+    @property
+    def vocab_size(self) -> int:
+        return len(self._vocab)
+
+    def get_vocab(self) -> Dict[str, int]:
+        return self._vocab.copy()
+
+    def _tokenize(self, text: str) -> List[str]:
+        """
+        Parses a string of moves into atomic tokens.
+        Input: "WPe2e4 BNg8f6"
+        Output: ["WP", "e2_f", "e4_t", "BN", "g8_f", "f6_t"]
+        """
+        if not text:
+            return []
+
+        tokens = []
+        raw_items = text.strip().split()
+        special_set = set(self.TOKENS_SPECIAL)
+
+        for item in raw_items:
+            # Pass through special tokens immediately
+            if item in special_set:
+                tokens.append(item)
+                continue
+
+            # Parse move structure
+            match = self.PATTERN_MOVE.match(item)
+            if not match:
+                tokens.append(self.unk_token)
+                continue
+
+            # Deconstruct parts
+            color, piece, src, dst, suffix = match.groups()
+
+            # 1. Actor (Who)
+            tokens.append(f"{color}{piece}")
+            
+            # 2. Origin (Where from)
+            tokens.append(f"{src}_f")
+            
+            # 3. Destination (Where to)
+            tokens.append(f"{dst}_t")
+
+            # 4. Promotion (Transformation)
+            # Check for suffixes like "=Q" or trailing chars
+            if suffix:
+                if "=" in suffix:
+                    # Look for the character immediately following '='
+                    eq_idx = suffix.find("=")
+                    if eq_idx + 1 < len(suffix):
+                        promo_char = suffix[eq_idx + 1].lower()
+                        if promo_char in self.CHARS_PROMO:
+                            tokens.append(promo_char)
+        
+        return tokens
+
+    def _convert_token_to_id(self, token: str) -> int:
+        return self._vocab.get(token, self.unk_token_id)
+
+    def _convert_id_to_token(self, index: int) -> str:
+        return self._id_to_token.get(index, self.unk_token)
+
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        """Joins tokens into a space-separated string, filtering out specials."""
+        special_set = set(self.TOKENS_SPECIAL)
+        return " ".join(t for t in tokens if t not in special_set)
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not os.path.exists(save_directory):
+            os.makedirs(save_directory, exist_ok=True)
+
+        filename = "vocab.json"
+        if filename_prefix:
+            filename = f"{filename_prefix}-{filename}"
+            
+        full_path = os.path.join(save_directory, filename)
+
+        with open(full_path, "w", encoding="utf-8") as f:
+            json.dump(self._vocab, f, indent=2, ensure_ascii=False)
+
+        return (full_path,)
+
+    @classmethod
+    def build_vocab_from_dataset(cls, *args: Any, **kwargs: Any) -> "ChessTokenizer":
+        """Compatibility method for training pipelines."""
+        return cls()

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

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0237f27c24fd35ae9075a49a691de6ee8b66113e3c207a0ec0847bcce221ccf5
+size 5777

vocab.json

@@ -0,0 +1,150 @@
+{
+  "[PAD]": 0,
+  "[BOS]": 1,
+  "[EOS]": 2,
+  "[UNK]": 3,
+  "WP": 4,
+  "WN": 5,
+  "WB": 6,
+  "WR": 7,
+  "WQ": 8,
+  "WK": 9,
+  "BP": 10,
+  "BN": 11,
+  "BB": 12,
+  "BR": 13,
+  "BQ": 14,
+  "BK": 15,
+  "a1_f": 16,
+  "b1_f": 17,
+  "c1_f": 18,
+  "d1_f": 19,
+  "e1_f": 20,
+  "f1_f": 21,
+  "g1_f": 22,
+  "h1_f": 23,
+  "a2_f": 24,
+  "b2_f": 25,
+  "c2_f": 26,
+  "d2_f": 27,
+  "e2_f": 28,
+  "f2_f": 29,
+  "g2_f": 30,
+  "h2_f": 31,
+  "a3_f": 32,
+  "b3_f": 33,
+  "c3_f": 34,
+  "d3_f": 35,
+  "e3_f": 36,
+  "f3_f": 37,
+  "g3_f": 38,
+  "h3_f": 39,
+  "a4_f": 40,
+  "b4_f": 41,
+  "c4_f": 42,
+  "d4_f": 43,
+  "e4_f": 44,
+  "f4_f": 45,
+  "g4_f": 46,
+  "h4_f": 47,
+  "a5_f": 48,
+  "b5_f": 49,
+  "c5_f": 50,
+  "d5_f": 51,
+  "e5_f": 52,
+  "f5_f": 53,
+  "g5_f": 54,
+  "h5_f": 55,
+  "a6_f": 56,
+  "b6_f": 57,
+  "c6_f": 58,
+  "d6_f": 59,
+  "e6_f": 60,
+  "f6_f": 61,
+  "g6_f": 62,
+  "h6_f": 63,
+  "a7_f": 64,
+  "b7_f": 65,
+  "c7_f": 66,
+  "d7_f": 67,
+  "e7_f": 68,
+  "f7_f": 69,
+  "g7_f": 70,
+  "h7_f": 71,
+  "a8_f": 72,
+  "b8_f": 73,
+  "c8_f": 74,
+  "d8_f": 75,
+  "e8_f": 76,
+  "f8_f": 77,
+  "g8_f": 78,
+  "h8_f": 79,
+  "a1_t": 80,
+  "b1_t": 81,
+  "c1_t": 82,
+  "d1_t": 83,
+  "e1_t": 84,
+  "f1_t": 85,
+  "g1_t": 86,
+  "h1_t": 87,
+  "a2_t": 88,
+  "b2_t": 89,
+  "c2_t": 90,
+  "d2_t": 91,
+  "e2_t": 92,
+  "f2_t": 93,
+  "g2_t": 94,
+  "h2_t": 95,
+  "a3_t": 96,
+  "b3_t": 97,
+  "c3_t": 98,
+  "d3_t": 99,
+  "e3_t": 100,
+  "f3_t": 101,
+  "g3_t": 102,
+  "h3_t": 103,
+  "a4_t": 104,
+  "b4_t": 105,
+  "c4_t": 106,
+  "d4_t": 107,
+  "e4_t": 108,
+  "f4_t": 109,
+  "g4_t": 110,
+  "h4_t": 111,
+  "a5_t": 112,
+  "b5_t": 113,
+  "c5_t": 114,
+  "d5_t": 115,
+  "e5_t": 116,
+  "f5_t": 117,
+  "g5_t": 118,
+  "h5_t": 119,
+  "a6_t": 120,
+  "b6_t": 121,
+  "c6_t": 122,
+  "d6_t": 123,
+  "e6_t": 124,
+  "f6_t": 125,
+  "g6_t": 126,
+  "h6_t": 127,
+  "a7_t": 128,
+  "b7_t": 129,
+  "c7_t": 130,
+  "d7_t": 131,
+  "e7_t": 132,
+  "f7_t": 133,
+  "g7_t": 134,
+  "h7_t": 135,
+  "a8_t": 136,
+  "b8_t": 137,
+  "c8_t": 138,
+  "d8_t": 139,
+  "e8_t": 140,
+  "f8_t": 141,
+  "g8_t": 142,
+  "h8_t": 143,
+  "b": 144,
+  "n": 145,
+  "q": 146,
+  "r": 147
+}