Math normalization: do not crash on invalid format (#331)

Wrap the asserts in a try except so that malformatted generations do not crash lighteval. See the original implem: https://github.com/hendrycks/math/blob/main/modeling/eval_math_gpt.py

Author: guipenedo

src/lighteval/metrics/normalizations.py

@@ -98,17 +98,20 @@ def _remove_boxed(text: str | None) -> str:
         """
         if text is None:
             return ""
-        if "\\boxed " in text:
-            left = "\\boxed "
-            assert text[: len(left)] == left
-            return text[len(left) :]
+        try:
+            if "\\boxed " in text:
+                left = "\\boxed "
+                assert text[: len(left)] == left
+                return text[len(left) :]
 
-        left = "\\boxed{"
+            left = "\\boxed{"
 
-        assert text[: len(left)] == left
-        assert text[-1] == "}"
+            assert text[: len(left)] == left
+            assert text[-1] == "}"
 
-        return text[len(left) : -1]
+            return text[len(left) : -1]
+        except Exception:
+            return ""
 
     def _last_boxed_only_string(text: str) -> str | None:
         """Extract the last \\boxed{...} or \\fbox{...} element from a string."""