arxiv BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Dataset