RIDE: Enhancing Large Language Model Alignment through Restyled In-Context Learning Demonstration Exemplars