RLTHF: Targeted Human Feedback for LLM Alignment