Question: Do we need a supervised dataset with X & Y pairs? #45
Description
one thing which is unclear, for prefix fine tuning do we always need labelled dataset? Say we want to do Q/A on proprietary data, would we need q/a as x/y pairs?You use webnlg data for summarisation, which does not seem to have summaries defined. Would you have any views on the kind of dataset needed for fine tuning with this method?