VLLM + Math-Verify fixes (#603)

* nits

* 🥰 pretty 🥰

---------

Co-authored-by: Hynek Kydlicek <[email protected]>

Author: hynky1999

src/lighteval/metrics/utils/extractive_match_utils.py

@@ -398,7 +398,10 @@ def extract_expr(match: re.Match, timeout_seconds: int) -> tuple[str | sympy.Exp
 
         decimal = decimal.replace(",", ".")
         number_str = f"{integer}{decimal}"
-        number = Number(number_str)
+        try:
+            number = Number(number_str)
+        except Exception:
+            return None, number_str
 
         if is_percentage:
             number = convert_to_pct(number)

src/lighteval/models/model_input.py

@@ -120,7 +120,11 @@ def to_vllm_dict(self) -> dict:
 
         # Task specific sampling params to set in model: n, best_of, use_beam_search
         # Generation specific params to set in model: logprobs, prompt_logprobs
-        return {sampling_params_to_vllm_naming.get(k, k): v for k, v in asdict(self).items() if v is not None}
+        x = {sampling_params_to_vllm_naming.get(k, k): v for k, v in asdict(self).items() if v is not None}
+        # VLLM max_tokens is 16 by default, however the pipeline expect the max_tokens to be None, if the user didn't specify it
+        if not x.get("max_tokens"):
+            x["max_tokens"] = None
+        return x
 
     def to_vllm_openai_dict(self) -> dict:
         """Selects relevant generation and sampling parameters for vllm and openai models.