explicit config modification

tasks/text.py

@@ -62,24 +62,38 @@ async def evaluate_text(request: TextEvaluationRequest):
         model_name = "Tonic/climate-guard-toxic-agent"
         tokenizer_name = "answerdotai/ModernBERT-base"
         
-        # Initialize config with specific parameters from config.json
+        # Load base config
         config = AutoConfig.from_pretrained(model_name)
-        config.architectures = ["ModernBertForSequenceClassification"]
-        config.model_type = "modernbert"
-        config.num_labels = 8
-        config.problem_type = "single_label_classification"
-        config.hidden_size = 768
-        config.num_attention_heads = 12
-        config.num_hidden_layers = 22
-        config.intermediate_size = 1152
-        config.max_position_embeddings = 8192
-        config.layer_norm_eps = 1e-05
-        config.classifier_dropout = 0.0
+        
+        # Remove problematic bias configurations
+        config_dict = config.to_dict()
+        bias_keys = ['attention_bias', 'classifier_bias', 'decoder_bias', 'mlp_bias', 'norm_bias']
+        for key in bias_keys:
+            if key in config_dict:
+                del config_dict[key]
+        
+        # Set essential configurations
+        config_dict.update({
+            "architectures": ["ModernBertForSequenceClassification"],
+            "model_type": "modernbert",
+            "num_labels": 8,
+            "problem_type": "single_label_classification",
+            "hidden_size": 768,
+            "num_attention_heads": 12,
+            "num_hidden_layers": 22,
+            "intermediate_size": 1152,
+            "max_position_embeddings": 8192,
+            "layer_norm_eps": 1e-05,
+            "classifier_dropout": 0.0
+        })
+        
+        # Create new config from cleaned dict
+        config = AutoConfig.from_dict(config_dict)
         
         # Load tokenizer
         tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
         
-        # Load model with specific configuration
+        # Load model with cleaned config
         model = AutoModelForSequenceClassification.from_pretrained(
             model_name,
             config=config,
@@ -124,14 +138,9 @@ async def evaluate_text(request: TextEvaluationRequest):
         predictions = []
         with torch.no_grad():
             for batch in test_loader:
-                # Move batch to device
                 batch = {k: v.to(device) for k, v in batch.items()}
-                
-                # Get model outputs
                 outputs = model(**batch)
                 preds = torch.argmax(outputs.logits, dim=-1)
-                
-                # Add batch predictions to list
                 predictions.extend(preds.cpu().numpy().tolist())
                 
         # Clean up GPU memory