How do you justify the use of hand-written exemplars in...

The argument for ‘unsupervised’ pipelines often ignores that they require at least one positive exemplar to function. Adding two or three additional hand-written examples is a one-time, fixed cost that significantly improves the quality of tens of thousands of generated pairs. To resolve the disagreement, propose a held-out pilot study to compare downstream retriever recall between vanilla-generated pairs and those generated using contrastive, hand-written exemplars.