[bugfix] guard _repair_ms_bench against an empty messages list

swift/dataset/dataset/llm.py

@@ -93,6 +93,10 @@ def preprocess(self, row):
 def _repair_ms_bench(messages: str) -> Optional[List[Dict[str, str]]]:
     if isinstance(messages, str):
         messages = ast.literal_eval(messages)
+    if not messages:
+        # A row with no messages can't be repaired; skip it like the MOSS case
+        # below instead of crashing the whole dataset load on messages[0].
+        return None
     default_system = 'You are a helpful assistant.'
     messages: List[Dict[str, str]]
     if messages[0]['from'] == 'system' and messages[0]['value'] == default_system:

tests/general/test_repair_ms_bench.py

@@ -0,0 +1,46 @@
+import unittest
+
+from swift.dataset.dataset.llm import _repair_ms_bench
+
+
+class TestRepairMsBench(unittest.TestCase):
+    """Pure unit tests for the ms_bench messages repair function (no network)."""
+
+    def test_empty_messages_returns_none(self):
+        # An empty row can't be repaired; it must be skipped (None) like the MOSS
+        # case rather than crashing the whole dataset load on messages[0].
+        self.assertIsNone(_repair_ms_bench('[]'))
+        self.assertIsNone(_repair_ms_bench([]))
+
+    def test_strips_default_system_message(self):
+        messages = [
+            {
+                'from': 'system',
+                'value': 'You are a helpful assistant.'
+            },
+            {
+                'from': 'user',
+                'value': 'hi'
+            },
+        ]
+        self.assertEqual(_repair_ms_bench(messages), [{'from': 'user', 'value': 'hi'}])
+
+    def test_keeps_a_normal_conversation(self):
+        messages = [
+            {
+                'from': 'user',
+                'value': 'hi'
+            },
+            {
+                'from': 'assistant',
+                'value': 'hello'
+            },
+        ]
+        self.assertEqual(_repair_ms_bench(messages), messages)
+
+    def test_skips_moss_rows(self):
+        self.assertIsNone(_repair_ms_bench([{'from': 'user', 'value': 'moss reply'}]))
+
+
+if __name__ == '__main__':
+    unittest.main()