What are the substantial changes from 3.0 to 3.1 (70B) in terms of training approach? They don't seem to say how the training data differed just that both were 15T. I gather 3.0 was just a preview run and 3.1 was distilled down from the 405B somehow.