Refine or train from scratch after updating dataset

How do you decide what to do after you’ve updated your dataset? Do you further refine your existing model or train from scratch. I assume it depends on how drastically you changed the data.