Merged Safe_dpo_helpful with base model for vLLM inference

initial commit