Although large language models (LLMs) have demonstrated strong instruction-following ability to be helpful, they are further supposed to be controlled and guided by rules in real-world scenarios to be safe, and accurate in responses. This demands the possession of →