Update modeling.py

modeling.py

@@ -11,7 +11,7 @@ from transformers import (
 
 
 # ============================================================
-# CONFIG CLASS
+# CONFIG CLASS (CRITICAL FIX HERE)
 # ============================================================
 class EmoAxisConfig(PretrainedConfig):
     model_type = "emoaxis"
@@ -21,15 +21,20 @@ class EmoAxisConfig(PretrainedConfig):
         self.num_labels = num_labels
         self.base_model_name = base_model_name
 
-    # 🚨 CRITICAL FIX: Overload from_pretrained to prevent argument clashes.
+    # 🚨 ULTIMATE FIX FOR THE TYPERROR 🚨
+    # We overload this method to manually filter out arguments that cause the clash.
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
-        # Filter arguments that cause TypeErrors in nested calls
+        # Filter the arguments that AutoConfig passes internally, causing the 'multiple values' error
         kwargs.pop("return_unused_kwargs", None)
         kwargs.pop("config", None)
         
-        # Call the parent method with cleaned arguments
-        return super().from_pretrained(pretrained_model_name_or_path, **kwargs)
+        # Call the base PretrainedConfig's method directly to avoid conflicts with super()
+        return PretrainedConfig.from_pretrained(
+            cls, # Pass the class itself (EmoAxisConfig) as the first argument
+            pretrained_model_name_or_path, 
+            **kwargs
+        )
 
 
 # ============================================================
@@ -74,11 +79,12 @@ class Classifier(nn.Module):
 class EmoAxis(PreTrainedModel):
     config_class = EmoAxisConfig
     
-    # FIX: Accept *args and **kwargs to avoid initialization errors.
+    # FIX: Accept *args and **kwargs for resilience
     def __init__(self, config, *args, **kwargs): 
         super().__init__(config, *args, **kwargs)
         
-        # FIX: Load base encoder from config to ensure fine-tuned weights are used.
+        # FIX: Load base encoder from its CONFIG to ensure fine-tuned weights are used 
+        # (prevents reinitialization of the base model weights)
         base_encoder = AutoModel.from_config(AutoConfig.from_pretrained(config.base_model_name))
         
         self.encoder = Encoder(base_encoder)