fix: correct output scores computation

app.py

@@ -49,11 +49,11 @@ def classify_pairs(model, tokenizer, prompts, responses):
     with torch.inference_mode():
         outputs = model(**input_ids)
         scores = torch.sigmoid(outputs.logits)
-        scores = [score for score in scores if float(score) > 0.5]
+        scores = [[score for score in s if float(score) > 0.5] for s in scores]
         
-        unsafety_score = sum(scores) / len(scores) if len(scores) > 0 else 0.0
+        unsafety_scores = [sum(s) / len(s) if len(s) > 0 else 0.0 for s in scores]
 
-    return unsafety_score
+    return unsafety_scores
 
 
 @spaces.GPU()