Do-Not-Answer: A Dataset for Evaluating Safeguards in LLMs