AFAIK (not much) it definitely helps to train on longer sequences even with rope/yarn and is needed if you care about long context performance (and not just the long context capability).