What is the minimum number of training samples needed to get a very good result with BART and PEGASUS. I’m training an headline generator, and 12000 samples does not seems to be enough.
Also, how many numbers of epochs is advised when fine-tuning these models.