They probably put most of the effort into the 65B model, the 7B model was just trained so they could get an idea of the scaling behaviour. It makes sense to use the same amount of training steps, then.