Wrap around any model to output differentially private prediction sets with finite sample validity on any dataset.