Comment by kazinator

Training a model isn't redistribution; only when you give someone a copy of the model can we think about there being a problem. At that point, you are not training, but redistributing a derived work.