dev

app.py

@@ -52,24 +52,22 @@ def fine_tune_model(model, dataset_name, hub_id, api_key, num_epochs, batch_size
    
     
         # Load the model and tokenizer
-             
-        
-        print(f"BATCH: {batch_size}")
+                 
         # Set training arguments
         training_args = TrainingArguments(
             output_dir='/data/results',
             eval_strategy="steps",  # Change this to steps
             save_strategy='steps',
-            learning_rate=lr*0.000001,
+            learning_rate=lr*0.00001,
             per_device_train_batch_size=int(batch_size),
             per_device_eval_batch_size=int(batch_size), 
             num_train_epochs=int(num_epochs),
             weight_decay=0.01,
-            gradient_accumulation_steps=int(grad),
+            #gradient_accumulation_steps=int(grad),
             #max_grad_norm = 3.0, 
             load_best_model_at_end=True,
-            #metric_for_best_model="accuracy",
-            #greater_is_better=True,
+            metric_for_best_model="accuracy",
+            greater_is_better=True,
             logging_dir='/data/logs',
             logging_steps=10,
             #push_to_hub=True,
@@ -230,7 +228,7 @@ def run_train(dataset_name, hub_id, api_key, num_epochs, batch_size, lr, grad):
         lora_dropout=0.1,  # Dropout for LoRA layers
         bias="none"  # Bias handling
     )
-    #model = get_peft_model(model, lora_config)
+    model = get_peft_model(model, lora_config)
     result = fine_tune_model(model, dataset_name, hub_id, api_key, num_epochs, batch_size, lr, grad)
     return result
 # Create Gradio interface