#multitask

Question 1: Single choice

Can a single fine-tuning minibatch for T5 mix examples from different tasks, for example summarization and translation?

Yes: every task is text-to-text, with a task description in the input, sharing one model and loss1
No: each task needs its own output layer, so a minibatch must come from one task2
No: the decoder's causal mask requires all targets in one minibatch to have exactly the same number of tokens3
Yes, but only if the tasks have the same number of output classes4