Update llm_clients.py

llm_clients.py

@@ -104,13 +104,12 @@ Rules:
 def judge_answers(answer1, answer2):
 
     judge_prompt = f"""
-You are an AI judge.
-
 Choose which answer is better.
 
 Reply ONLY with:
-1  (if answer1 is better)
-2  (if answer2 is better)
+1
+or
+2
 
 Answer 1:
 {answer1}
@@ -122,9 +121,8 @@ Answer 2:
     result = call_llama(
         [{"role": "user", "content": judge_prompt}],
         temperature=0
-    )
+    ).strip()
 
-    if "2" in result:
+    if result == "2":
         return 2
     return 1
-