FWIW this is extractive question answering so the output of the model can only be a span of the input paragraph.

SQuAD is the prototypical example of a dataset for this task, see https://rajpurkar.github.io/SQuAD-explorer/