SD2's dataset was deduplicated prior to training. That's why the paper is about SD1, which was a prototype model completed before Stability even had VC raises.