Knowledge Unlearning for Large Language Models

📄arXiv • 🤗HFPaper • 🤗HF Collection

This repository provides the official PyTorch implementation of our paper:

ReLearn: Unlearning via Learning for Large Language Models

Haoming Xu¹, Ningyuan Zhao², Liming Yang³, Sendong Zhao⁴, Shumin Deng⁵, Mengru Wang¹, Bryan Hooi⁵, Nay Oo⁵, Huajun Chen¹, Ningyu Zhang¹

¹Zhejiang University, ²Xiamen University, ³Tsinghua University, ⁴Harbin Institute of Technology, ⁵National University of Singapore

🎉 News

🏆 Our team won 2nd place in the SEMEval 2025 Challenge on Unlearning Sensitive Content from Large Language Models! Check out our implementation in the Semeval25 directory.

🌟 Overview

📦 Installation

# Create and activate conda environment
conda create -n relearn python=3.10.15
conda activate relearn

# Install PyTorch with CUDA support
conda install pytorch pytorch-cuda=11.8 -c pytorch -c nvidia
conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit

# Install dependencies
pip install -r requirements.txt
pip install flash-attn --no-build-isolation

🚀 Quick Start

1. Data Augmentation

cd dataAugument
bash augu.sh

2. Model Training

Currently supports:

• Llama3-8b instruct
• Gemma2-2b-it
• Llama2-7b chat

cd baselines/pretrain_scripts/
bash kud-pt.sh

3. Unlearning Process

cd baselines/unlearn_scripts/
bash kud-relearn.sh

4. Evaluation

cd evals
bash merge_all.sh
bash inf_all.sh
bash eval_all.sh

🔧 Supported Methods

Method	Script	Source & Notes
GA / NPO	unlearn/baselines/unlearn_scripts/kud-baselines.sh	Source
SURE	unlearn/baselines/unlearn_scripts/kud-baselines.sh	Source
Memflex	unlearn/baselines/unlearn_scripts/kud-baselines.sh	Source (Iterative version)
ReLearn	unlearn/baselines/unlearn_scripts/kud-relearn.sh	Our Method
ReLearn_dpo	unlearn/baselines/unlearn_scripts/kud-relearn.sh	Uses augmented data as positive samples

📂 Open Resources

Pretrained Models

• Llama-2-7b-chat-KnowUnDo-Privacy (Vanilla)
  🔗 ModelScope

• Llama-2-7b-chat-TOFU-Forget10-ReLearn
  🔗 Google Drive

• Llama-2-7b-chat-KnowUnDo-Privacy-ReLearn
  🔗 Google Drive

Datasets

• Augmented KnowUnDo Privacy Dataset
  🔗 Google Drive

🙏 Acknowledgements

This repository builds upon code from:TOFU and MUSE.

We thank the authors for their excellent work.

📝 Citation

If you find this work useful for your research, please cite our paper:

@misc{xu2025relearnunlearninglearninglarge,
      title={ReLearn: Unlearning via Learning for Large Language Models}, 
      author={Haoming Xu and Ningyuan Zhao and Liming Yang and Sendong Zhao and 
              Shumin Deng and Mengru Wang and Bryan Hooi and Nay Oo and 
              Huajun Chen and Ningyu Zhang},
      year={2025},
      eprint={2502.11190},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.11190}
}