fix msbench bug (#749)

Author: hjh0119

swift/llm/utils/dataset.py

@@ -559,7 +559,8 @@ def _repair_agent_conversations(conversations: str,
 
 
 def _repair_ms_bench(conversations: str) -> Dict[str, str]:
-    conversations = ast.literal_eval(conversations)
+    if isinstance(conversations, str):
+        conversations = ast.literal_eval(conversations)
     default_system = 'You are a helpful assistant.'
     if conversations[0]['from'] == 'system' and conversations[0][
             'value'] == default_system: