These
inputs should be used for sequence to sequence tasks, such as translation or summarization, and are usually built in a
way specific to each model.