Upload consensus_fusion.py with huggingface_hub

consensus_fusion.py

@@ -101,7 +101,12 @@ def normalize_answers(answers: List[str]) -> Dict[str, List[int]]:
             key_clean = _normalize_answer(key)
             
             if MATH_VERIFY_AVAILABLE:
-                if verify(clean, key_clean, parsing_timeout=None):
+                try:
+                    match = verify(clean, key_clean)
+                except (ValueError, Exception):
+                    # Fallback when math_verify fails in threaded env (signal.alarm restriction)
+                    match = (str(clean) == str(key_clean))
+                if match:
                     normalized_groups[key].append(idx)
                     matched = True
                     break