Relearn: Machine Unlearning Framework for LLMs

Overview

Relearn is a comprehensive framework for machine unlearning in large language models. The project addresses the critical need for removing specific knowledge from trained models while maintaining overall performance—essential for privacy compliance, model refinement, and ethical AI deployment.

Key Features

• Dual-phase Unlearning: Combines targeted knowledge removal with performance preservation
• Scalable Architecture: Applicable to various LLM architectures (GPT, BERT, T5, etc.)
• Benchmark Suite: Comprehensive evaluation metrics for unlearning effectiveness

Technical Highlights

• Developed novel learning-based unlearning methods that reduce unwanted knowledge retention by 60%+
• Maintained task performance within 2% of original models
• Published at top-tier venues with open-source implementation

Impact

This work has implications for:

• Privacy-preserving machine learning
• GDPR and data protection compliance
• Model refinement and knowledge management
• Responsible AI deployment

Publications

• Relearn: Unlearning via Learning for Large Language Models (2025)
• ZJUKLAB at SemEval-2025 Task 4: Unlearning via Model Merging (2025)