If you find this repo to be useful, please consider cite our paper. Thank you.

    @inproceedings{
    zhang2025catastrophic,
    title={Catastrophic Failure of {LLM} Unlearning via Quantization},
    author={Zhiwei Zhang and Fali Wang and Xiaomin Li and Zongyu Wu and Xianfeng Tang and Hui Liu and Qi He and Wenpeng Yin and Suhang Wang},
    booktitle={The Thirteenth International Conference on Learning Representations},
    year={2025},
    url={https://openreview.net/forum?id=lHSeDYamnz}
    }

To create a conda environment for Python 3.10, run:

conda env create -f environment.yml
conda activate py310

Two corpora News and Books and the associated target models are available as follows:

Before proceeding, load all the data from HuggingFace to the root of this repostiory by running the following instruction:

python load_data.py

To unlearn the target model using our baseline method including SURE, run unlearn.py in the baselines folder. Example scripts baselines/scripts/unlearn_news.sh and scripts/unlearn_books.sh in the baselines folder demonstrate the usage of unlearn.py. Here is an example:

algo="ga"
CORPUS="news"

python unlearn.py \
        --algo $algo \
        --model_dir $TARGET_DIR --tokenizer_dir 'meta-llama/Llama-2-7b-hf' \
        --data_file $FORGET --retain_data_file $RETAIN \
        --out_dir "./ckpt/$CORPUS/$algo" \
        --max_len $MAX_LEN --epochs $EPOCHS --lr $LR \
        --alpha 1 --threshold 90 \
        --per_device_batch_size $PER_DEVICE_BATCH_SIZE

• algo: Unlearning algorithm to run (ga, ga_gdr, ga_klr, npo, npo_gdr, npo_klr, ga_gdr_sure, ga_klr_sure, npo_gdr_sure, npo_klr_sure, rmu).
• model_dir: Directory of the target model.
• tokenizer_dir: Directory of the tokenizer.
• data_file: Forget set.
• retain_data_file: Retain set for GDR/KLR regularizations if required by the algorithm.
• out_dir: Directory to save the unlearned model.
• max_len: Maximum input length (default: 2048).
• per_device_batch_size, epochs, lr: Hyperparameters.
• alpha: weight for utility constraint.
• threshold: threshold to filter out salient modules (e.g. 90).

Resulting models are saved in the ckpt folder as shown:

ckpt
├── news/
│   ├── ga/
│   │   ├── checkpoint-102
│   │   ├── checkpoint-204
│   │   ├── checkpoint-306
│   │   └── ...
│   └── npo/
│       └── ...
└── books/
    ├── ga
    └── ...

To evaluate your unlearned model(s), run eval.py from the root of this repository with the following command-line arguments:

• --model_dirs: A list of directories containing the unlearned models. These can be either HuggingFace model directories or local storage paths.
• --names: A unique name assigned to each unlearned model in --model_dirs. The length of --names should match the length of --model_dirs.
• --corpus: The corpus to use for evaluation. Options are news or books.
• --out_file: The name of the output file. The file will be in CSV format, with each row corresponding to an unlearning method from --model_dirs, and columns representing the metrics specified by --metrics.
• --tokenizer_dir (Optional): The directory of the tokenizer. Defaults to meta-llama/Llama-2-7b-hf, which is the default tokenizer for LLaMA.
• --metrics (Optional): The metrics to evaluate. Options are verbmem_f (VerbMem Forget), privleak (PrivLeak), knowmem_f (KnowMem Forget), and knowmem_r (Knowmem Retain, i.e., Utility). Defaults to evaluating all these metrics.
• --temp_dir (Optional): The directory for saving intermediate computations. Defaults to temp.
• --quantize_4bit (Optional): whether quantize the model to 4 bit.
• --quantize_8bit (Optional): whether quantize the model to 8 bit.
• set quantize_4bit=0, quantize_8bit=0 to test model in full precision. quantize_4bit=1, quantize_8bit=0 to test model in 4-bit. quantize_4bit=0, quantize_8bit=1 to test model in 8-bit

Example scripts eval.sh in the root of this repository demonstrate the usage of eval.py.

Run the following command with placeholder values:

python eval.py \
  --model_dirs "data/model1" "data/model2" \
  --names "model1" "model2" \
  --corpus books \
  --out_file "out.csv"

Our code is mainly built on muse_bench. The evaluation code on task Gen, Tru, Fac and Flu is based on RWKU. We appreciate their open-source code.