LLM-NEO: Parameter Efficient Knowledge Distillation for Large Language Models