Split, Unlearn, Merge: Leveraging Data Attributes for More Effective Unlearning in LLMs