README.md

MEOW

This is the official implementation of the paper "MEOW: Memory Supervised LLM Unlearning via inverted facts"

Step1: Invert Facts

You can use inverted facts to ToFU-1%, ToFU-5%, and ToFU-10% provided in ./data/tofu/generated_data, or run python -m data.invert to generate new inverted facts, and in ./data/convert.py, we provide some implementation for data preprocessing for your reference.

Step2: Memory Supervised

Under the folder memo, we provide three methods for quantifying the memorization in LLMs, called MA, EL and MEMO (proposed in our paper). After run one of them, you can get data with memo in ./data/tofu/memo_data. Similar to the Invert Facts step, you can use provided memo data.

Here is a simple example:

python -m memo.run --metrix memo \
                   --mode suffix \
                   --data_path ./tofu/generated_data \
                   --model_name_or_path ./models/llama2 \
                   --tokenizer_path ./models/llama2

Step3: LLM Unlearning

In this step, we will start to LLM Unlearning.

First of all, you can download the origin model from locuslab/ToFU.

Secondly, run bash scripts/forget.sh to unlearn. We put all the hyperparameters in ./meow/tofu/config/forget.yaml.

At last, run bash eval.sh and you can find your eval result in the folder eval_results. We put all the hyperparameters in ./meow/tofu/config/eval.yaml

Experiments on NLU and NLG Benchmarks

In this step, we will conduct experiments on NLU and NLG Benchmarks using unlearned LLM.

For NLU evaluation, here is a simple start script:

python -m evaluate.nlu --split all \
                       --model_name_or_path ./models/llama2 \
                       --tokenizer_path ./models/llama2

For NLG evaluation, here is a simple start script:

python -m evaluate.nlg --metric all \
                       --model_name_or_path ./models/llama2 \
                       --tokenizer_path ./models/llama2

Additional Features

We implement a robot to remind you of the completion of experiments based on Feishu in ./meow/message.py. You can just remove them from codes if you don't need it. 😉

Cite Us

@misc{gu2024meowmemorysupervisedllm,
      title={MEOW: MEMOry Supervised LLM Unlearning Via Inverted Facts}, 
      author={Tianle Gu and Kexin Huang and Ruilin Luo and Yuanqi Yao and Yujiu Yang and Yan Teng and Yingchun Wang},
      year={2024},
      eprint={2409.11844},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2409.11844}, 
}