This is the official github page for the paper Evaluating Deep Unlearning in Large Language Model. This repository provides the code for using the benchmark EDU-RELAT, the evaluation code given any unlearning results, and the scripts of running four unlearning methods presented in the paper.

1. Install the environment

conda env create -f environment.yml
conda activate unlearning
pip install flash-attn --no-build-isolation

2. Install lm-evaluation-harness for general benchmark evaluation

git clone --depth 1 https://github.com/EleutherAI/lm-evaluation-harness
cp plugin_for_lm_eval/* lm-evaluation-harness/lm_eval/models/
cd lm-evaluation-harness
pip install -e .

3. Download the model checkpoints from this link that is finetuned on our synthetic data EDU-RELAT. The layout would be

deep_unlearning/
    ft_model_checkpoint/
        ft_gpt2-xl
        ft_llama2-7b
        ft_llama3-8b
        ft_phi

3. Set up your huggingface key HF_TOKEN in the os environment.

Here we provide the code to load the knowledge base and rule sets

1. Load the QA-form knowledge base as huggingfact datasets

from datasets import Dataset
dataset_relationships = Dataset.from_dict(torch.load("synthetic_data/family_relationships.pt")) #load the facts in family relationships
dataset_biographies = Dataset.from_dict(torch.load("synthetic_data/family_biographies.pt")) #load the facts in biographies

The question and answer are with the key question4 and answer4.

2. Load the rule set

from calculate_recall_and_acc import Rule
rule_list = torch.load("synthetic_data/family_rule.pt")

3. Load the relationships as a list of tuple

from calculate_recall_and_acc import Person
(edge_list,relation_list, _, _) = torch.load("synthetic_data/family-200-graph.pt") #edge_list is a list of pairs of two people; relation_list is a list of relationthips in string, e.g. child.

We select a random subset of size 55 from the facts in family relationship to evaluate the deep unlearning. Given any unlearn_target_data_id in 0-54, the id of fact in relationships is

shuffled_edge_id_list = torch.load("synthetic_data/subsample.pt")
shuffled_unlearn_data_id = shuffled_edge_id_list[unlearn_target_data_id]

For any unlearning method, suppose relationships_correct.pt and biographies_correct.pt are two 0-1 vectors saved under the directory input_dir, which indicate the retained facts in family relationships and biographies after the unlearning the fact unlearn_target_data_id. The following script will calculate the recall and accuracy for this unlearning method to unlearn the target data.

python calculate_recall_and_acc.py --unlearn_data_id $unlearn_target_data_id --input_dir $input_dir

We provide example_for_evaluation/relationships_correct.pt and example_for_evaluation/biographies_correct.pt as an example for calculating the recall and accuracy, by running python calculate_recall_and_acc.py --unlearn_data_id 1 --input_dir example_for_evaluation.

In the paper, we tested with four unlearning methods: gradient ascent (GA), Negative preference optimization (NPO), task vector (TV), who's harry potter (WHP). The hyperparameter list of each method is saved in config/model_config.yaml. The related scripts are saved in ./unlearning_methods. By set any unlearning_method (ga, npo, tv, whp), any target_model (phi, gpt2-xl, llama2-7b, llama3-8b), and unlearn_target_data_id (0-54), the script is

bash unlearning_methods/${unlearning_methods}.sh $target_model $unlearn_target_data_id

After running unlearning methods, the code will save two 0-1 vectors relationships_correct.pt and biographies_correct.pt under the directory scripts_unlearning_checkpoint/${unlearning_methods}/${target_model}/${unlearn_target_data_id}/checkpoint-${hyperparameter}. Then run the script in the above section to calcualte the recall and accuracy.

If you find our codebase and dataset beneficial, please cite our work:

@misc{wu2024evaluatingdeepunlearninglarge,
      title={Evaluating Deep Unlearning in Large Language Models}, 
      author={Ruihan Wu and Chhavi Yadav and Russ Salakhutdinov and Kamalika Chaudhuri},
      year={2024},
      eprint={2410.15153},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.15153}, 
}

We would like to thank the authors of TOFU and MUSE. Our code is built upon the github repository of them.