I don't know a single person who would prefer to change a little bit of the source dataset and then spend $100k training from scratch. Everyone finetunes existing models not simply because the source dataset is not available (for some models, it is available), but because retraining from scratch sucks for modifying models.