This research introduces a general framework that improves existing fine-tuning-based unlearning methods for large language models while maintaining their utility for normal tasks.

• Addresses the challenge of removing copyrighted and privacy-sensitive data from LLMs
• Enhances both gradient ascent-based and suppression-based unlearning approaches
• Preserves model performance on standard tasks while effectively removing target information

This work is crucial for security practitioners as it provides a practical approach to selectively remove sensitive data from deployed LLMs without compromising their overall performance, helping organizations meet compliance requirements while maintaining valuable AI assets.

A General Framework to Enhance Fine-tuning-based LLM Unlearning