At the very least, I think transformers benefit from progressive increase in sample length. But building a principled curriculum based on abstract semantic-leve

At the very least, I think transformers benefit from progressive increase in sample length. But building a principled curriculum based on abstract semantic-level properties of the content doesn't seem to work, or we don't know how how prioritize it.