Removes default bert scorer init (#234)

* move init of bertscorer to avoid downloading model by default

* Update src/lighteval/metrics/metrics_sample.py

Co-authored-by: Nathan Habib <[email protected]>

---------

Co-authored-by: Nathan Habib <[email protected]>

src/lighteval/metrics/metrics_sample.py

@@ -424,9 +424,7 @@ def __init__(
             normalize_pred (callable, optional): Function to use to normalize the predicted strings.
                 Defaults to None if no normalization is applied.
         """
-        self.bert_scorer = BERTScorer(
-            model_type="microsoft/deberta-large-mnli", lang="en", rescale_with_baseline=True, num_layers=9
-        )
+        self.bert_scorer = None
 
         self.normalize_gold = normalize_gold
         self.normalize_pred = normalize_pred
@@ -441,6 +439,12 @@ def compute(self, golds: list[str], predictions: list[str]) -> dict:
         Returns:
             dict: Scores over the current sample's items.
         """
+        if self.bert_scorer is None:
+            hlog_warn("The first metric computation step might be a bit longer as we need to download the model.")
+            # We only initialize on first compute
+            self.bert_scorer = BERTScorer(
+                model_type="microsoft/deberta-large-mnli", lang="en", rescale_with_baseline=True, num_layers=9
+            )
         golds = as_list(golds)
         predictions = as_list(predictions)
         # Normalize