Hello, I would like to fine-tune RoBERTa using a triplet loss function, following what was described in the SBERT paper (sec. 3). What is the best way for building a fine-tuning dataset? My plan is to modify the Trainer class.
Hi @anon98739735 . Did you find a way to do this ? I have a similar problem in hand and looking for ways to fine-tune RoBERTa using instances of triplets. Any help would be appreciated. Thanks