If you choose this second option, there are three possibilities you can use to gather all the input Tensors RoBERTa has almost similar architecture as compare to BERT, but in order to improve the results on BERT architecture, the authors made some simple design changes in its archi