DP2Unlearning

Paper: DP2Unlearning: An Efficient and Guaranteed Unlearning Framework for Llms

Installation

To set up the environment for the project, create a conda environment using the following command:

$ conda create --name torch-env pytorch torchvision pytorch-cuda=12.1 -c pytorch -c nvidia
$ conda activate torch-env

Then, install the following libraries:

pip install datasets accelerate evaluate matplotlib hydra-core omegaconf peft rouge_score tqdm einops packaging bitsandbytes scipy ninja

Also you may install additional libraries if required.

Traditional Retraining from Scratch (Benchmark retain model)

To perform traditional retraining from scratch, run the following command:

python finetune.py --config-path /home/user_name/project_name/config --config-name finetune.yaml

Do necessary modification in finetune.yaml file based on your hardware and GPU capacity.

Unlearning Ready Training (Disclosure protected base model)

To train a disclosure-protected base model for unlearning, use one of the following options:

python DP2U-MLM.py %(to transform raw data to disclosure protected data using DP-MLM)
python Train_dp_MLM.py --config-path /home/user_name/project_name/config --config-name Train_dp_MLM.yaml

or

python Train_dp_SGD.py --config-path /home/user_name/project_name/config --config-name Train_dp_SGD.yaml

Do necessary modification in Train_dp_MLM.yaml or Train_dp_SGD.yaml based on your hardware and GPU capacity.

DP2Unlearning Fine-Tuning

For DP2Unlearning fine-tuning, run:

python FT_BaseModel.py --config-path /home/user_name/project_name/config --config-name FT_BaseModel.yaml

Do necessary modification to FT_BaseModel.yaml based on forgetting percentage (1%:retain99, 5%:retain95, or 10%:retain90)

Approximate Unlearning Baselines Fine-Tuning

To perform approximate unlearning fine-tuning, execute the following:

python forget.py --config-path /home/user_name/project_name/config --config-name forget.yaml

Evaluation

To evaluate the models, use this command:

python evaluate_util.py --config-path /home/user_name/project_name/config --config-name eval_everything.yaml

You need to provide the specific model path that you wish to evaluate.

Aggregation

To aggregate the evaluation statistics, use:

python aggregate_eval_stat.py --config-path /home/user_name/project_name/config --config-name aggregate_eval_stat.yaml

Ensure you have the paths to your results:

retain_result=${path_to_traditional_retraining_from_scratch}
ckpt_result=${path_to_your_unlearned_method}

Beyond KS Test

To run the Beyond KS Test, execute:

python Beyond_KS_test.py --config-path /home/user_name/project_name/config --config-name aggregate_eval_stat.yaml

---

The baseline methods are implemented from [1]