Chess Challenge submission by theophile-ltt

README.md

@@ -0,0 +1,26 @@
+---
+library_name: transformers
+tags:
+- chess
+- llm-course
+- chess-challenge
+license: mit
+---
+
+# chess-theophile
+
+Chess model submitted to the LLM Course Chess Challenge.
+
+## Submission Info
+
+- **Submitted by**: [theophile-ltt](https://huggingface.co/theophile-ltt)
+- **Parameters**: 705,280
+- **Organization**: LLM-course
+
+## Model Details
+
+- **Architecture**: Chess Transformer (GPT-style)
+- **Vocab size**: 84
+- **Embedding dim**: 128
+- **Layers**: 4
+- **Heads**: 4

config.json

@@ -0,0 +1,24 @@
+{
+  "architectures": [
+    "ChessForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "model.ChessConfig",
+    "AutoModelForCausalLM": "model.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": 4,
+  "pad_token_id": 0,
+  "tie_weights": true,
+  "transformers_version": "4.57.1",
+  "vocab_size": 84
+}

model.py

@@ -0,0 +1,348 @@
+"""
+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 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 = 84,
+        n_embd: int = 128,
+        n_layer: int = 7,
+        n_head: int = 4,
+        n_ctx: int = 512,
+        n_inner: Optional[int] = 256,
+        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)
+    
+    def forward(
+        self,
+        x: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, seq_len, _ = x.size()
+        
+        qkv = self.c_attn(x)
+        q, k, v = qkv.split(self.n_embd, dim=2)
+        
+        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)
+        
+        attn_weights = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
+        
+        causal = torch.ones(seq_len, seq_len, device=x.device, dtype=torch.bool).tril()
+        attn_weights = attn_weights.masked_fill(~causal.view(1, 1, seq_len, seq_len), float("-inf"))
+        
+        if attention_mask is not None:
+            attention_mask = attention_mask.to(torch.bool)
+            attn_weights = attn_weights.masked_fill(~attention_mask.view(batch_size, 1, 1, seq_len), float("-inf"))
+        
+        attn_weights = F.softmax(attn_weights, dim=-1)
+        attn_weights = self.dropout(attn_weights)
+        
+        attn_output = torch.matmul(attn_weights, v)
+        
+        attn_output = attn_output.transpose(1, 2).contiguous().view(
+            batch_size, seq_len, self.n_embd
+        )
+        
+        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:
+        x = x + self.attn(self.ln_1(x), attention_mask=attention_mask)
+        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)
+        
+        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)
+        
+        self.h = nn.ModuleList([
+            TransformerBlock(config) for _ in range(config.n_layer)
+        ])
+        
+        self.ln_f = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+        
+        if config.tie_weights:
+            self._tied_weights_keys = ["lm_head.weight"]
+        
+        self.post_init()
+        
+        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):
+        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
+        
+        if position_ids is None:
+            position_ids = torch.arange(seq_len, device=device).unsqueeze(0).expand(batch_size, -1)
+        
+        token_embeds = self.wte(input_ids)
+        position_embeds = self.wpe(position_ids)
+        hidden_states = self.drop(token_embeds + position_embeds)
+        
+        for block in self.h:
+            hidden_states = block(hidden_states, attention_mask=attention_mask)
+        
+        hidden_states = self.ln_f(hidden_states)
+        
+        logits = self.lm_head(hidden_states)
+        
+        loss = None
+        if labels is not None:
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            
+            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,
+        )
+
+
+# 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:e3feb7d1cdd8b7853baf23f364986e78b5bdcf707f9537fb243b5a7e5fdd2185
+size 2825520

special_tokens_map.json

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

tokenizer.py

@@ -0,0 +1,218 @@
+"""
+Custom Chess Tokenizer for the Chess Challenge.
+
+This tokenizer treats each move as a sequence of structured tokens derived from 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
+import re
+from typing import Dict, List, Optional, Sequence, Union
+
+from transformers import PreTrainedTokenizer
+
+
+_MOVE_RE = re.compile(
+    r"^(?P<side>[WB])"
+    r"(?P<piece>[PNBRQK])"
+    r"(?P<src>[a-h][1-8])"
+    r"(?P<dst>[a-h][1-8])"
+    r"(?P<rest>.*)$"
+)
+
+
+class ChessTokenizer(PreTrainedTokenizer):
+    """
+    A structured tokenizer for chess moves.
+
+    Each move is decomposed into:
+      SIDE_(W/B), PIECE_(P/N/B/R/Q/K), SQ_<src>, SQ_<dst>,
+      and optional flags: CAPTURE, CHECK, MATE, CASTLE, PROMO_(Q/R/B/N).
+
+    This avoids UNK explosions when using a move-as-token vocabulary.
+    """
+
+    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]"
+
+    # Fixed token set
+    SIDE_W = "SIDE_W"
+    SIDE_B = "SIDE_B"
+
+    PIECES = ["P", "N", "B", "R", "Q", "K"]
+
+    PROMO_PREFIX = "PROMO_"
+    CAPTURE = "CAPTURE"
+    CHECK = "CHECK"
+    MATE = "MATE"
+    CASTLE = "CASTLE"
+
+    def __init__(
+        self,
+        vocab_file: Optional[str] = None,
+        vocab: Optional[Dict[str, int]] = None,
+        **kwargs,
+    ):
+        kwargs.pop("pad_token", None)
+        kwargs.pop("bos_token", None)
+        kwargs.pop("eos_token", None)
+        kwargs.pop("unk_token", None)
+
+        self._pad_token = self.PAD_TOKEN
+        self._bos_token = self.BOS_TOKEN
+        self._eos_token = self.EOS_TOKEN
+        self._unk_token = self.UNK_TOKEN
+
+        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._build_fixed_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,
+        )
+
+    def _build_fixed_vocab(self) -> Dict[str, int]:
+        special = [self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN]
+        sides = [self.SIDE_W, self.SIDE_B]
+        pieces = [f"PIECE_{p}" for p in self.PIECES]
+        squares = [f"SQ_{file}{rank}" for file in "abcdefgh" for rank in "12345678"]
+        promos = [f"{self.PROMO_PREFIX}{p}" for p in ["Q", "R", "B", "N"]]
+        flags = [self.CAPTURE, self.CHECK, self.MATE, self.CASTLE]
+        tokens = special + sides + pieces + squares + promos + flags
+        return {tok: i for i, tok in enumerate(tokens)}
+
+    @classmethod
+    def build_vocab_from_dataset(cls, *args, **kwargs) -> "ChessTokenizer":
+        """
+        Kept for API compatibility with the template training script.
+        This tokenizer uses a fixed vocabulary (no dataset-dependent pruning).
+        """
+        return cls()
+
+    @classmethod
+    def build_vocab_from_iterator(cls, *args, **kwargs) -> "ChessTokenizer":
+        """
+        Kept for API compatibility. This tokenizer uses a fixed vocabulary.
+        """
+        return cls()
+
+    @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]:
+        tokens: List[str] = []
+        moves = text.strip().split()
+        for mv in moves:
+            tokens.extend(self._tokenize_move(mv))
+        return tokens
+
+    def _tokenize_move(self, move: str) -> List[str]:
+        m = _MOVE_RE.match(move)
+        if not m:
+            return [self.UNK_TOKEN]
+
+        side = m.group("side")
+        piece = m.group("piece")
+        src = m.group("src")
+        dst = m.group("dst")
+        rest = m.group("rest") or ""
+
+        out: List[str] = []
+        out.append(self.SIDE_W if side == "W" else self.SIDE_B)
+        out.append(f"PIECE_{piece}")
+        out.append(f"SQ_{src}")
+        out.append(f"SQ_{dst}")
+
+        promo = self._parse_promotion(rest)
+        if promo is not None:
+            out.append(f"{self.PROMO_PREFIX}{promo}")
+
+        if "(x)" in rest or "x" in rest:
+            out.append(self.CAPTURE)
+
+        if "(+*)" in rest or "++" in rest or "#" in rest:
+            out.append(self.MATE)
+        elif "(+)" in rest or "+" in rest:
+            out.append(self.CHECK)
+
+        if "(o)" in rest or "(O)" in rest or "O-O" in rest:
+            out.append(self.CASTLE)
+
+        return out
+
+    def _parse_promotion(self, rest: str) -> Optional[str]:
+        m = re.search(r"=([QRBNqrbn])", rest)
+        if m:
+            return m.group(1).upper()
+        return None
+
+    def _convert_token_to_id(self, token: str) -> int:
+        return self._vocab.get(token, self._vocab[self.UNK_TOKEN])
+
+    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}
+        out: List[str] = []
+        for t in tokens:
+            if t in special:
+                continue
+            out.append(t)
+        return " ".join(out)
+
+    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,)
+
+    def decode(self, token_ids: Union[int, Sequence[int]], skip_special_tokens: bool = False, **kwargs) -> str:
+        if isinstance(token_ids, int):
+            ids = [token_ids]
+        elif "torch" in str(type(token_ids)):
+            ids = token_ids.detach().cpu().flatten().tolist()
+        else:
+            ids = list(token_ids)
+
+        toks = [self._convert_id_to_token(i) for i in ids]
+        if skip_special_tokens:
+            special = {self.PAD_TOKEN, self.BOS_TOKEN, self.EOS_TOKEN, self.UNK_TOKEN}
+            toks = [t for t in toks if t not in special]
+        return self.convert_tokens_to_string(toks)

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

vocab.json

@@ -0,0 +1,86 @@
+{
+  "[PAD]": 0,
+  "[BOS]": 1,
+  "[EOS]": 2,
+  "[UNK]": 3,
+  "SIDE_W": 4,
+  "SIDE_B": 5,
+  "PIECE_P": 6,
+  "PIECE_N": 7,
+  "PIECE_B": 8,
+  "PIECE_R": 9,
+  "PIECE_Q": 10,
+  "PIECE_K": 11,
+  "SQ_a1": 12,
+  "SQ_a2": 13,
+  "SQ_a3": 14,
+  "SQ_a4": 15,
+  "SQ_a5": 16,
+  "SQ_a6": 17,
+  "SQ_a7": 18,
+  "SQ_a8": 19,
+  "SQ_b1": 20,
+  "SQ_b2": 21,
+  "SQ_b3": 22,
+  "SQ_b4": 23,
+  "SQ_b5": 24,
+  "SQ_b6": 25,
+  "SQ_b7": 26,
+  "SQ_b8": 27,
+  "SQ_c1": 28,
+  "SQ_c2": 29,
+  "SQ_c3": 30,
+  "SQ_c4": 31,
+  "SQ_c5": 32,
+  "SQ_c6": 33,
+  "SQ_c7": 34,
+  "SQ_c8": 35,
+  "SQ_d1": 36,
+  "SQ_d2": 37,
+  "SQ_d3": 38,
+  "SQ_d4": 39,
+  "SQ_d5": 40,
+  "SQ_d6": 41,
+  "SQ_d7": 42,
+  "SQ_d8": 43,
+  "SQ_e1": 44,
+  "SQ_e2": 45,
+  "SQ_e3": 46,
+  "SQ_e4": 47,
+  "SQ_e5": 48,
+  "SQ_e6": 49,
+  "SQ_e7": 50,
+  "SQ_e8": 51,
+  "SQ_f1": 52,
+  "SQ_f2": 53,
+  "SQ_f3": 54,
+  "SQ_f4": 55,
+  "SQ_f5": 56,
+  "SQ_f6": 57,
+  "SQ_f7": 58,
+  "SQ_f8": 59,
+  "SQ_g1": 60,
+  "SQ_g2": 61,
+  "SQ_g3": 62,
+  "SQ_g4": 63,
+  "SQ_g5": 64,
+  "SQ_g6": 65,
+  "SQ_g7": 66,
+  "SQ_g8": 67,
+  "SQ_h1": 68,
+  "SQ_h2": 69,
+  "SQ_h3": 70,
+  "SQ_h4": 71,
+  "SQ_h5": 72,
+  "SQ_h6": 73,
+  "SQ_h7": 74,
+  "SQ_h8": 75,
+  "PROMO_Q": 76,
+  "PROMO_R": 77,
+  "PROMO_B": 78,
+  "PROMO_N": 79,
+  "CAPTURE": 80,
+  "CHECK": 81,
+  "MATE": 82,
+  "CASTLE": 83
+}