Update modeling_xlm_roberta.py

modeling_xlm_roberta.py

@@ -499,7 +499,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
                 Overwrites any setting from convert_to_numpy
             device(`torch.device`, *optional*, defaults to None):
                 Which torch.device to use for the computation
-            normalize_embeddings(`bool`, *optional*, defaults to False):
+            normalize_embeddings(`bool`, *optional*, defaults to True):
                 If set to true, returned vectors will have length 1. In that case, the
                 faster dot-product (util.dot_score) instead of cosine similarity can
                 be used.