Opt-Out: Investigating Entity-Level Unlearning for Large Language Models via Optimal Transport

Source code for our ACL 2025 paper Opt-Out: Investigating Entity-Level Unlearning for Large Language Models via Optimal Transport.

This codebase implements various unlearning methods to make language models "forget" specific entities while preserving their general capabilities.

🔧 Installation

To install requirements:

conda create -n optout python=3.12.9
conda activate optout
pip install -r requirements.txt

🤗 Data

We provide the ELUDe (Entity-Level Unlearning Dataset) on Hugging Face: https://huggingface.co/datasets/6rightjade/ELUDe

ELUDe is a comprehensive machine unlearning dataset focused on the removal of entire entities from large language models (LLMs). The dataset includes:

• 20 real-world target entities (the entities listed below)
• 144 unique neighboring entities from Wikipedia

📊 Available Entities

The codebase supports unlearning for 20 different entities:

1. Donald_Trump
2. Elizabeth_II
3. Barack_Obama
4. Cristiano_Ronaldo
5. Michael_Jackson
6. Elon_Musk
7. Lady_Gaga
8. Adolf_Hitler
9. Eminem
10. Lionel_Messi
11. Justin_Bieber
12. Freddie_Mercury
13. Kim_Kardashian
14. Johnny_Depp
15. Steve_Jobs
16. Dwayne_Johnson
17. Michael_Jordan
18. Taylor_Swift
19. Stephen_Hawking
20. Kanye_West

🧠 Unlearning Methods

Core Methods

• original - The original performance of the model
• icu - In-Context Unlearning: Prompting baseline (Guardrail)
• ga - Gradient Ascent: Uses gradient ascent for unlearning
• dpo - Direct Preference Optimization: Uses DPO for unlearning
• npo - Negative Preference Optimization: Uses NPO for unlearning
• idk - I Don't Know: Makes the model respond with "I don't know"

Data Augmentation Options

You can combine core methods with the following modifiers (except original and icu):

• +rt - Retain Data: Includes neighboring entity data to preserve nearby knowledge
• +wd - World Data: Uses Alpaca GPT-4 data for maintaining general knowledge (we use Alpaca GPT-4 data from here)
• +ot - Optimal Transport: Adds Wasserstein regularization for better unlearning

Example Method Combinations

• npo+rt+wd+ot - NPO with retain data, world data, and optimal transport (Opt-Out)
• dpo+rt+wd - DPO with retain and world data
• ga+rt - Gradient ascent with retain data only
• idk+wd - IDK method with world data only

🚀 Usage

Training

Use the training script to fine-tune models for entity unlearning:

bash scripts/train.sh

Evaluation

Run evaluation on trained models:

bash scripts/eval.sh

📁 Directory Structure

Opt-Out/
├── run.py              # Main training/evaluation script
├── trainer.py          # Custom trainer implementation
├── model.py            # Model loading utilities  
├── dataset.py          # Data loading and processing
├── evaluator.py        # Evaluation logic
├── scripts/            # Execution scripts
│   ├── train.sh        # Training script
│   └── eval.sh         # Evaluation script
├── data/               # External data

📚 Citation

If you use this codebase, please cite our paper:

@article{choi2025optout,
  title={Opt-Out: Investigating Entity-Level Unlearning for Large Language Models via Optimal Transport},
  author={Choi, Minseok and Rim, Daniel and Lee, Dohyun and Choo, Jaegul},
  journal={arXiv preprint arXiv:2406.12329},
  year={2025}
}