From 0d28cb42e1d69bbee2797a4fc10548f32c658563 Mon Sep 17 00:00:00 2001
From: Yufeng He <40085740+he-yufeng@users.noreply.github.com>
Date: Sat, 20 Jun 2026 23:29:27 +0800
Subject: [PATCH] [bugfix] guard _repair_ms_bench against an empty messages
 list

_repair_ms_bench reads messages[0] right after ast.literal_eval, so a row whose
messages are empty (e.g. the string "[]") raises IndexError and aborts the whole
ms_bench dataset load instead of just skipping that row. The function already
returns None to skip rows it can't use (the MOSS case), so do the same for an
empty list. Added pure unit tests for the empty, default-system, normal and
MOSS cases.
---
 swift/dataset/dataset/llm.py          |  4 +++
 tests/general/test_repair_ms_bench.py | 46 +++++++++++++++++++++++++++
 2 files changed, 50 insertions(+)
 create mode 100644 tests/general/test_repair_ms_bench.py

diff --git a/swift/dataset/dataset/llm.py b/swift/dataset/dataset/llm.py
index a144964f22..55169b752e 100644
--- a/swift/dataset/dataset/llm.py
+++ b/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:
diff --git a/tests/general/test_repair_ms_bench.py b/tests/general/test_repair_ms_bench.py
new file mode 100644
index 0000000000..d9e10f490b
--- /dev/null
+++ b/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()