return scores per question and not summary

score.py

@@ -62,7 +62,7 @@ client = OpenAI(
 )
 
 
-def calculate_gpt4o_score(queries, user_answers, correct_answers):
+def calculate_gpt4o_scores(queries, user_answers, correct_answers):
     """
     Calculate the score for the Icelandic Wiki QA benchmark.
     """
@@ -87,6 +87,7 @@ def calculate_gpt4o_score(queries, user_answers, correct_answers):
         score = score_filter.search(chat).group(1).lower()
         scores.append(score)
 
-    return sum(
-        1 if score == "excellent" else 0.5 if score == "fair" else 0 for score in scores
-    ) / len(scores)
+    scores_numeric = [
+        1.0 if score == "excellent" else 0.5 if score == "fair" else 0.0 for score in scores
+    ]
+    return scores_numeric