test edit

semf1.py

@@ -339,6 +339,7 @@ class SemF1(evaluate.Metric):
             gpu: DEVICE_TYPE = False,
             batch_size: int = 32,
             verbose: bool = False,
+            aggregate: bool = True,
     ) -> List[Scores]:
         """
             Compute precision, recall, and F1 scores for given predictions and references.
@@ -358,6 +359,7 @@ class SemF1(evaluate.Metric):
             :param gpu: GPU device to use.
             :param batch_size: Batch size for encoding.
             :param verbose: Flag to indicate verbose output.
+            :param aggregate: Flag to determine if output should be averaged
 
             :return: List of Scores dataclass with precision, recall, and F1 scores.
         """
@@ -422,3 +424,4 @@ class SemF1(evaluate.Metric):
 
         return results
 
+# TEST