We propose LUNAR, an novel LLM unlearning method by redirecting the representation of unlearned data for effective, controlled and efficient unlearning.

Open-source framework for LLM unlearning supporting multiple benchmarks and methods.