Safety-Tuned LLaMAs: Lessons From Improving the Safety of Large Language Models that Follow Instructions