I had the same question. I feel like they should clarify this. But also I don't see how they would want to apply that restriction in this case, since using their models to generate training data will make them more money and makes fine-tuning vastly more manageable.