The base model wouldn't do that though, it would just predict the most likely follow up, which could e.g. simply be more instructions. After instruction fine-tuning the model does no longer "predict" tokens in this way.