Which Retain Set Matters for LLM Unlearning? A Case Study on Entity Unlearning