Fix evaluate crash on tasks with no stop sequences

mlx_lm/evaluate.py

@@ -35,7 +35,10 @@ def _rstrip_until(s, untils):
     l = len(s)
     f = [s.find(u) for u in untils]
     f = [l if x < 0 else x for x in f]
-    return s[: min(f)]
+    # A task may legitimately specify no stop sequences -- ifeval sets
+    # `until: []` because the completion is meant to run to max_gen_toks --
+    # and there is then nothing to truncate at.
+    return s[: min(f, default=l)]
 
 
 def _lstrip(s, pattern):

tests/test_evaluate.py

@@ -5,7 +5,7 @@
 
 import mlx.core as mx
 
-from mlx_lm.evaluate import MLXLM
+from mlx_lm.evaluate import MLXLM, _rstrip_until
 
 
 class TestMLXLM(unittest.TestCase):
@@ -55,5 +55,25 @@ def mock_score_fn(batch):
         self.assertEqual(len(call_args_list[2][0][0]), 1)  # Third batch: 1 item
 
 
+class TestRstripUntil(unittest.TestCase):
+    def test_truncates_at_first_stop_sequence(self):
+        self.assertEqual(_rstrip_until("abc<eos>def", ["<eos>"]), "abc")
+
+    def test_truncates_at_the_earliest_of_several(self):
+        self.assertEqual(_rstrip_until("abc\n\nQ:", ["Q:", "\n\n"]), "abc")
+
+    def test_returns_the_whole_string_when_nothing_matches(self):
+        self.assertEqual(_rstrip_until("abc", ["<eos>"]), "abc")
+
+    def test_no_stop_sequences(self):
+        # Tasks may legitimately specify none: lm-eval's ifeval (and every one
+        # of its multilingual variants) sets `until: []` so the completion runs
+        # to max_gen_toks. There is then nothing to truncate at.
+        self.assertEqual(_rstrip_until("abc", []), "abc")
+
+    def test_no_stop_sequences_on_empty_string(self):
+        self.assertEqual(_rstrip_until("", []), "")
+
+
 if __name__ == "__main__":
     unittest.main()