Chess Challenge submission by Oceane28

README.md

@@ -0,0 +1,31 @@
+---
+library_name: transformers
+tags:
+- chess
+- llm-course
+- chess-challenge
+license: mit
+---
+
+# chess-6-Oceane28
+
+Chess model submitted to the LLM Course Chess Challenge.
+
+## Submission Info
+
+- **Submitted by**: [Oceane28](https://huggingface.co/Oceane28)
+- **Parameters**: 996,428
+- **Organization**: Oceane28
+
+## Usage
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model = AutoModelForCausalLM.from_pretrained("Oceane28/chess-6-Oceane28", trust_remote_code=True)
+tokenizer = AutoTokenizer.from_pretrained("Oceane28/chess-6-Oceane28", 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": 8,
+  "n_inner": 354,
+  "n_layer": 6,
+  "pad_token_id": 0,
+  "tie_weights": true,
+  "transformers_version": "4.57.5",
+  "vocab_size": 136,
+  "auto_map": {
+    "AutoConfig": "model.ChessConfig",
+    "AutoModelForCausalLM": "model.ChessForCausalLM"
+  }
+}

model.py

@@ -0,0 +1,396 @@
+"""
+Modèle Transformer pour le Chess Challenge (1M paramètres).
+
+Ce module fournit une architecture transformer de style GPT conçue
+pour respecter la contrainte stricte de moins de 1 million de paramètres.
+
+Composants clés :
+- ChessConfig : Configuration des hyperparamètres.
+- ChessForCausalLM : Le modèle principal pour la prédiction du prochain coup.
+"""
+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, AutoConfig, AutoModelForCausalLM
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+# -----------------------------------------------------------------------------
+# Configuration
+# -----------------------------------------------------------------------------
+
+class ChessConfig(PretrainedConfig):
+    """
+    Classe de configuration pour le modèle Chess Transformer.
+    
+    Conçue pour un budget de paramètres très serré (< 1M).
+    
+    Répartition du budget (avec les valeurs par défaut de ton ami) :
+    - Vocabulaire (Embeddings) : 72 * 92 = ~6,6k
+    - Embeddings de position : 256 * 92 = ~23,5k
+    - Couches Transformer : 11 couches * (~85k par couche) = ~935k
+    - Tête LM (liée aux embeddings) : 0 paramètre supplémentaire
+    - Total : ~970k paramètres (juste en dessous de 1M).
+    """
+    
+    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
+        self.dropout = dropout
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.tie_weights = tie_weights
+        self.tie_word_embeddings = bool(tie_weights)
+
+
+# -----------------------------------------------------------------------------
+# Modules du Transformer
+# -----------------------------------------------------------------------------
+
+class MultiHeadAttention(nn.Module):
+    """
+    Module d'attention multi-têtes standard.
+    Inclut le masquage causal pour empêcher le modèle de "voir le futur".
+    """
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__()
+        
+        assert config.n_embd % config.n_head == 0, \
+            f"n_embd ({config.n_embd}) doit être divisible par 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
+        
+        # Projection combinée Q, K, V pour l'efficacité
+        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)
+        
+        # Masque causal (registre persistent=False pour ne pas le sauvegarder dans le checkpoint)
+        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()
+        
+        # Calcul de Q, K, V
+        qkv = self.c_attn(x)
+        q, k, v = qkv.split(self.n_embd, dim=2)
+        
+        # Remodelage pour l'attention multi-têtes
+        # (batch, seq_len, n_head, head_dim) -> (batch, n_head, seq_len, head_dim)
+        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)
+        
+        # Attention produit scalaire (Scaled Dot-Product Attention)
+        attn_weights = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
+        
+        # Application du masque causal (le futur est masqué avec -inf)
+        causal_mask = self.bias[:, :, :seq_len, :seq_len]
+        attn_weights = attn_weights.masked_fill(causal_mask == 0, float("-inf"))
+        
+        # Application du masque d'attention (pour le 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)
+        
+        # Application de l'attention aux valeurs
+        attn_output = torch.matmul(attn_weights, v)
+        
+        # Remise en forme
+        attn_output = attn_output.transpose(1, 2).contiguous().view(
+            batch_size, seq_len, self.n_embd
+        )
+        
+        # Projection de sortie
+        attn_output = self.c_proj(attn_output)
+        
+        return attn_output
+
+
+class FeedForward(nn.Module):
+    """
+    Réseau de neurones Feed-Forward (MLP).
+    Deux couches linéaires avec une activation GELU entre les deux.
+    """
+    
+    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):
+    """
+    Un bloc transformer unique contenant Attention et Feed-Forward.
+    Utilise la "Pre-normalization" (LayerNorm avant l'attention/FFN) pour la stabilité.
+    """
+    
+    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:
+        # Connexion résiduelle + Pre-norm Attention
+        x = x + self.attn(self.ln_1(x), attention_mask=attention_mask)
+        # Connexion résiduelle + Pre-norm FFN
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+# -----------------------------------------------------------------------------
+# Modèle Principal
+# -----------------------------------------------------------------------------
+
+class ChessForCausalLM(PreTrainedModel):
+    """
+    Modèle final pour la prédiction de coups (Causal Language Modeling).
+    
+    Architecture :
+    1. Embeddings (Tokens + Position)
+    2. Empilement de blocs Transformer
+    3. Tête linéaire finale (Projection vers le vocabulaire)
+    """
+    
+    config_class = ChessConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    
+    # Ignore l'avertissement de clé manquante car lm_head partage les poids avec wte
+    keys_to_ignore_on_load_missing = ["lm_head.weight"]
+    
+    def __init__(self, config: ChessConfig):
+        super().__init__(config)
+        
+        # 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)
+        
+        # Blocs Transformer
+        self.h = nn.ModuleList([
+            TransformerBlock(config) for _ in range(config.n_layer)
+        ])
+        
+        # LayerNorm final
+        self.ln_f = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        
+        # Tête de sortie (sans biais pour économiser des paramètres)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+        
+        # Gestion du partage de poids (Weight Tying)
+        if config.tie_weights:
+            self._tied_weights_keys = ["lm_head.weight"]
+        
+        # Initialisation des poids
+        self.post_init()
+        
+        # Forcer le lien des poids si configuré
+        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):
+        """Lie les poids de l'embedding d'entrée et de la tête de sortie."""
+        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):
+        """Initialisation des poids style GPT-2."""
+        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]:
+        """
+        Passe avant (Forward pass).
+        Calcule les logits et, si des étiquettes (labels) sont fournies, la perte (loss).
+        """
+        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
+        
+        # Création des IDs de position s'ils ne sont pas fournis
+        if position_ids is None:
+            position_ids = torch.arange(seq_len, device=device).unsqueeze(0).expand(batch_size, -1)
+        
+        # Calcul des embeddings (Token + Position)
+        token_embeds = self.wte(input_ids)
+        position_embeds = self.wpe(position_ids)
+        hidden_states = self.drop(token_embeds + position_embeds)
+        
+        # Passage dans les blocs Transformer
+        for block in self.h:
+            hidden_states = block(hidden_states, attention_mask=attention_mask)
+        
+        # Normalisation finale
+        hidden_states = self.ln_f(hidden_states)
+        
+        # Calcul des logits (prédiction du prochain token)
+        logits = self.lm_head(hidden_states)
+        
+        # Calcul de la perte (Training Loss)
+        loss = None
+        if labels is not None:
+            # On décale les logits et les labels d'un cran
+            # (le modèle doit prédire le token t+1 à partir du token t)
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            
+            # Perte CrossEntropy standard
+            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:
+        """
+        Génère le prochain coup pour une séquence donnée.
+        Utilisé pour l'inférence en jeu réel.
+        """
+        self.eval()
+        
+        # Récupère les logits pour la dernière position uniquement
+        outputs = self(input_ids)
+        logits = outputs.logits[:, -1, :] / temperature
+        
+        # Filtrage Top-K
+        if top_k is not None:
+            indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+            logits[indices_to_remove] = float("-inf")
+        
+        # Filtrage Top-P (Nucleus Sampling)
+        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)
+            
+            # On retire les tokens qui sont au-dessus du seuil cumulatif
+            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")
+        
+        # Échantillonnage final
+        probs = F.softmax(logits, dim=-1)
+        next_token = torch.multinomial(probs, num_samples=1)
+        
+        return next_token.item()
+
+# Enregistrement pour chargement automatique via AutoModel
+AutoConfig.register("chess_transformer", ChessConfig)
+AutoModelForCausalLM.register(ChessConfig, ChessForCausalLM)

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:50286b26484c3225b11c798752b682a11247037f5c023480787bef0680f51116
+size 3992160

special_tokens_map.json

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

tokenizer.py

@@ -0,0 +1,154 @@
+"""
+Role-marked square tokenizer for Chess Challenge.
+Each move is represented as: <from_square>_f <to_square>_t [promo?] [EOS]
+
+Examples:
+  WPe2e4        -> e2_f e4_t [EOS]
+  BPe7e8=Q      -> e7_f e8_t q [EOS]
+"""
+from __future__ import annotations
+import json
+import os
+import re
+from typing import Dict, List, Optional, Tuple, Any
+from transformers import PreTrainedTokenizer
+
+_MOVE_RE = re.compile(r"^([WB])([PNBRQK])([a-h][1-8])([a-h][1-8])(.*)$")
+_PROMO_RE = re.compile(r"=([QRBNqrbn])")
+
+SQUARES = [f"{f}{r}" for r in "12345678" for f in "abcdefgh"]
+PROMOS = ["q", "r", "b", "n"]
+
+class ChessTokenizer(PreTrainedTokenizer):
+    model_input_names = ["input_ids", "attention_mask"]
+    vocab_files_names = {"vocab_file": "vocab.json"}
+    
+    PAD_TOKEN = "[PAD]"
+    BOS_TOKEN = "[BOS]"
+    EOS_TOKEN = "[EOS]"
+    UNK_TOKEN = "[UNK]"
+
+    def __init__(
+        self,
+        vocab_file: Optional[str] = None,
+        vocab: Optional[Dict[str, int]] = None,
+        **kwargs: Any,
+    ):
+        self._pad_token = self.PAD_TOKEN
+        self._bos_token = self.BOS_TOKEN
+        self._eos_token = self.EOS_TOKEN
+        self._unk_token = self.UNK_TOKEN
+        
+        for k in ["pad_token", "bos_token", "eos_token", "unk_token"]:
+            kwargs.pop(k, None)
+            
+        if vocab is not None:
+            self._vocab = vocab
+        elif vocab_file is not None 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._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,
+        )
+
+    @staticmethod
+    def _build_fixed_vocab() -> Dict[str, int]:
+        tokens: List[str] = [
+            ChessTokenizer.PAD_TOKEN,
+            ChessTokenizer.BOS_TOKEN,
+            ChessTokenizer.EOS_TOKEN,
+            ChessTokenizer.UNK_TOKEN,
+        ]
+        tokens += [f"{sq}_f" for sq in SQUARES]
+        tokens += [f"{sq}_t" for sq in SQUARES]
+        tokens += PROMOS
+        return {tok: i for i, tok in enumerate(tokens)}
+
+    @property
+    def vocab_size(self) -> int:
+        return len(self._vocab)
+
+    def get_vocab(self) -> Dict[str, int]:
+        return dict(self._vocab)
+
+    @classmethod
+    def build_vocab_from_dataset(cls, *args: Any, **kwargs: Any) -> "ChessTokenizer":
+        return cls()
+
+    @classmethod
+    def build_vocab_from_iterator(cls, *args: Any, **kwargs: Any) -> "ChessTokenizer":
+        return cls()
+
+    def _tokenize(self, text: str) -> List[str]:
+        """Tokenize a space-separated list of dataset moves into role-marked tokens."""
+        text = (text or "").strip()
+        if not text:
+            return []
+            
+        out: List[str] = []
+        for move in text.split():
+            # Allow already-tokenized text (debugging)
+            if move in self._vocab:
+                out.append(move)
+                continue
+                
+            # Try to match Standard Lichess Format (WPe2e4)
+            m = _MOVE_RE.match(move)
+            if not m:
+                # If it's plain UCI like e2e4 or e7e8q
+                if re.fullmatch(r"[a-h][1-8][a-h][1-8][qrbn]?", move):
+                    src, dst = move[:2], move[2:4]
+                    out.append(f"{src}_f")
+                    out.append(f"{dst}_t")
+                    if len(move) == 5:
+                        out.append(move[4])
+                    out.append(self.EOS_TOKEN)
+                    continue
+                
+                # Unknown token
+                out.append(self.UNK_TOKEN)
+                out.append(self.EOS_TOKEN)
+                continue
+            
+            # Extract parts from WPe2e4...
+            _side, _piece, src, dst, suffix = m.groups()
+            out.append(f"{src}_f")
+            out.append(f"{dst}_t")
+            
+            promo = None
+            pm = _PROMO_RE.search(suffix or "")
+            if pm:
+                promo = pm.group(1).lower()
+            
+            if promo in PROMOS:
+                out.append(promo)
+                
+            out.append(self.EOS_TOKEN)
+            
+        return out
+
+    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:
+        return " ".join(tokens)
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        os.makedirs(save_directory, exist_ok=True)
+        name = "vocab.json" if not filename_prefix else f"{filename_prefix}-vocab.json"
+        path = os.path.join(save_directory, name)
+        with open(path, "w", encoding="utf-8") as f:
+            json.dump(self._vocab, f, indent=2, ensure_ascii=False)
+        return (path,)

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]",
+  "clean_up_tokenization_spaces": false,
+  "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:d8bf96856bc64ce4893cdd6ee8c82a58af1a49726bbbfb085d0b6a2d7c115ffd
+size 5777

vocab.json

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