Chess Challenge submission by MaximeMuhlethaler

model.py

@@ -24,7 +24,6 @@ class ChessConfig(PretrainedConfig):
         dropout=0.1,
         layer_norm_epsilon=1e-5,
         tie_weights=True,
-        # Valeurs par défaut strictes
         pad_token_id=0,
         bos_token_id=1,
         eos_token_id=2,
@@ -41,7 +40,7 @@ class ChessConfig(PretrainedConfig):
         self.layer_norm_epsilon = layer_norm_epsilon
         self.tie_weights = tie_weights
         
-        # On passe les IDs vitaux à kwargs pour le parent
+
         kwargs["pad_token_id"] = pad_token_id
         kwargs["bos_token_id"] = bos_token_id
         kwargs["eos_token_id"] = eos_token_id
@@ -118,7 +117,7 @@ class ChessForCausalLM(PreTrainedModel):
     def set_output_embeddings(self, new_embeddings): self.lm_head = new_embeddings
 
     def forward(self, input_ids, attention_mask=None, position_ids=None, labels=None, return_dict=None, **kwargs):
-        # 1. FIX TYPE RETOUR
+
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
         if return_dict is None: return_dict = True
 
@@ -132,17 +131,13 @@ class ChessForCausalLM(PreTrainedModel):
         x = self.ln_f(x)
         logits = self.lm_head(x)
 
-        # ---------------------------------------------------------
-        # 2. PATCH NUCLÉAIRE : On bannit 0, 1, 2, 3 en dur
-        # ---------------------------------------------------------
+
         if labels is None:
-            # PAD=0, BOS=1, EOS=2, UNK=3 (Les standards de ton tokenizer)
+
             nuclear_bad_ids = [0, 1, 2, 3] 
             
-            # On met -infini (impossible à choisir)
-            # Le slicing [:, :, ids] couvre tout le batch et toute la séquence
             logits[:, :, nuclear_bad_ids] = float("-inf")
-        # ---------------------------------------------------------
+
 
         loss = None
         if labels is not None:

tokenizer.py

@@ -108,17 +108,14 @@ class ChessTokenizer(PreTrainedTokenizer):
         from datasets import load_dataset
         from collections import Counter
         
-        # On charge en streaming pour aller vite
         ds = load_dataset(dataset_name, split="train", streaming=True)
-        ds = ds.take(50000) # 50k parties suffisent pour voir tous les coups possibles
+        ds = ds.take(50000) 
         
         counter = Counter()
         for ex in ds:
-            # On normalise avant de compter !
             moves = [normalize_move(t) for t in ex["text"].split()]
             counter.update(moves)
             
-        # On garde les tokens spéciaux + les N plus fréquents
         special = [cls.PAD_TOKEN, cls.BOS_TOKEN, cls.EOS_TOKEN, cls.UNK_TOKEN]
         most_common = counter.most_common(max_vocab_size - len(special))