fix df model

app.py

@@ -65,7 +65,7 @@ def response(num_responses, model, correct, prompt_ids):
     #     num_responses = len(responses)
 
     # sample num_responses for each model
-    responses = responses.groupby('model').apply(lambda x: x.sample(num_responses) if num_responses < len(x) else x).reset_index()
+    responses = responses.groupby('model').apply(lambda x: x.sample(num_responses) if num_responses < len(x) else x)
     return responses[['model', 'prompt', 'model_response', 'correct']]