Indeed - there’s a very good argument for using synthetic data to introduce diversity as long as you can avoid model collapse.