Update common.py

common.py

@@ -623,13 +623,13 @@ def resolve_single_judgment_dict(
     """Return the correct single answer grading judge."""
     if multi_turn:
         if question["category"] in NEED_REF_CATS:
-            return model_judgments_math[("gpt-4", "single-math-v1-multi-turn")]
-        return model_judgments_normal[("gpt-4", "single-v1-multi-turn")]
+            return model_judgments_math[("gpt-4o-2024-05-13", "single-math-v1-multi-turn")]
+        return model_judgments_normal[("gpt-4o-2024-05-13", "single-v1-multi-turn")]
 
     if question["category"] in NEED_REF_CATS:
-        return model_judgments_math[("gpt-4", "single-math-v1")]
+        return model_judgments_math[("gpt-4o-2024-05-13", "single-math-v1")]
     else:
-        return model_judgments_normal[("gpt-4", "single-v1")]
+        return model_judgments_normal[("gpt-4o-2024-05-13", "single-v1")]
 
 
 def get_pairwise_judge_explanation(gamekey, judgment_dict):