Graph

ReLearn: Learning new things for Large Language Models

Summary

The paper “ReLearn: Unlearning via Learning for Large Language Models” presents a novel method for unlearning in LLMs while preserving fluency and relevance. It introduces a data augmentation and fine-tuning pipeline as an alternative to ‘gradient ascent (GA)’ and ’negative preference optimization (NPO]’, which degrade linguistic coherence.

How to Implement This Paper

To implement ReLearn, we will follow these key steps:

1️⃣ Understanding the Core Approach

  • Data Augmentation: Generate diverse question-answer (QA) variations for forgetting while ensuring non-sensitive yet relevant responses.
  • Fine-Tuning: Replace the knowledge to be forgotten with relevant but non-sensitive responses.
  • Evaluation Metrics: Use the paper’s Knowledge Forgetting Rate (KFR), Knowledge Retention Rate (KRR), and Linguistic Score (LS) for performance assessment.

2️⃣ Setting Up the Development Environment

We need: