QA system

[mzuzic]

Hi,

thanks for your good work!

I'm testing your model (BigBirdRobertaQA) in a retriever-reader architecture. I retrieve the top 5 paragraphs (they are a couple of sentences each), merge them (joining them with a space between) and feed them as the context to the model.

1. Is that the best way to approach this? If so, what is the recommended max length for the tokenizer?
2. I expect not, but is there a way to influence what kind of answer will you get? (short, long, yes/no)