Why do we pass both question and answer as input to the model during training?

[ruiyigan]

I may be wrong here but from what I traced from their train.py file, it seems like they are training the model by passing both question + answer as the input and then evaluating the model's output with the answer. (Their question + answer is basically input_ids and labels is just the answer)

I abit curious as to why they did this (or I may be wrong) as I thought usually models are trainined by giving a question:answer pair rather than a (question + answer):answer pair.

Appreciate any help on this! Thank you