RWKU: Benchmarking Real-World Knowledge Unlearning for Large Language Models

🏠 Homepage | 📜 Paper | 🤗 Dataset | 🚀 Installation

News

2024-09-26: 🚀🚀 Our paper has been accepted at NeurIPS D&B Track 2024.
2024-06-18: We released our paper titled "RWKU: Benchmarking Real-World Knowledge Unlearning for Large Language Models".
2024-06-05: We released our dataset on the Huggingface.

Description

RWKU is a real-world knowledge unlearning benchmark specifically designed for large language models (LLMs). This benchmark contains 200 real-world unlearning targets and 13,131 multi-level forget probes, including 3,268 fill-in-the-blank probes, 2,879 question-answer probes, and 6,984 adversarial-attack probes. RWKU is designed based on the following three key factors:

For the task setting, we consider a more practical and challenging setting, similar to zero-shot knowledge unlearning. We provide only the unlearning target and the original model, without offering any forget corpus or retain corpus. In this way, it avoids secondary information leakage caused by the forget corpus and is not affected by the distribution bias of the retain corpus.
For the knowledge source, we choose real-world famous people from Wikipedia as the unlearning targets and demonstrate that such popular knowledge is widely present in various LLMs through memorization quantification, making it more suitable for knowledge unlearning. Additionally, choosing entities as unlearning targets can well clearly define the unlearning boundaries.
For the evaluation framework, we carefully design the forget set and the retain set to evaluate the model's capabilities from multiple real-world applications.
- Regarding the forget set, we evaluate the efficacy of knowledge unlearning at both the knowledge memorization (fill-in-the-blank style) and knowledge manipulation (question-answer style) abilities. Specifically, we also evaluate these two abilities through adversarial attacks to induce forgotten knowledge in the model. We adopt four membership inference attack (MIA) methods for knowledge memorization on our collected MIA set. We meticulously designed nine types of adversarial-attack probes for knowledge manipulation, including prefix injection, affirmative suffix, role playing, reverse query, and others.
- Regarding the retain set, we design a neighbor set to test the impact of neighbor perturbation, specifically focusing on the locality of unlearning. In addition, we assess the model utility on various downstream capabilities, including general ability, reasoning ability, truthfulness, factuality, and fluency.

Installation

git clone https://github.com/jinzhuoran/RWKU.git
conda create -n rwku python=3.10
conda activate rwku
cd RWKU
pip install -r requirements.txt

Dataset Download and Processing

One way is to load the dataset from Huggingface and preprocess it.

cd process
python data_process.py

Another way is to download the processed dataset directly from Google Drive.

cd LLaMA-Factory/data
bash download.sh

Unlearning Target

RWKU includes 200 famous people from The Most Famous All-time People Rank, such as Stephen King, Warren Buffett, Taylor Swift, etc. We demonstrate that such popular knowledge is widely present in various LLMs through memorization quantification, making it more suitable for unlearning.

from datasets import load_dataset
forget_target = load_dataset("jinzhuoran/RWKU", 'forget_target')['train'] # 200 unlearning targets

Evaluation Framework

RWKU mainly consists of four subsets, including forget set, neighbor set, MIA set and utility set. Evaluation Framework.

Forget Set

from datasets import load_dataset
forget_level1 = load_dataset("jinzhuoran/RWKU", 'forget_level1')['test'] # forget knowledge memorization probes
forget_level2 = load_dataset("jinzhuoran/RWKU", 'forget_level2')['test'] # forget knowledge manipulation probes
forget_level3 = load_dataset("jinzhuoran/RWKU", 'forget_level3')['test'] # forget adversarial attack probes

Neighbor Set

from datasets import load_dataset
neighbor_level1 = load_dataset("jinzhuoran/RWKU", 'neighbor_level1')['test'] # neighbor knowledge memorization probes
neighbor_level2 = load_dataset("jinzhuoran/RWKU", 'neighbor_level2')['test'] # neighbor knowledge manipulation probes

MIA Set

from datasets import load_dataset
mia_forget = load_dataset("jinzhuoran/RWKU", 'mia_forget') # forget member set
mia_retain = load_dataset("jinzhuoran/RWKU", 'mia_retain') # retain member set

Utility Set

from datasets import load_dataset
utility_general = load_dataset("jinzhuoran/RWKU", 'utility_general') # general ability
utility_reason = load_dataset("jinzhuoran/RWKU", 'utility_reason') # reasoning ability
utility_truthfulness = load_dataset("jinzhuoran/RWKU", 'utility_truthfulness') # truthfulness
utility_factuality = load_dataset("jinzhuoran/RWKU", 'utility_factuality') # factuality
utility_fluency = load_dataset("jinzhuoran/RWKU", 'utility_fluency') # fluency

Supported Unlearning Methods

In-Context Unlearning (ICU): We use specific instructions to make the model behave as if it has forgotten the target knowledge, without actually modifying the model parameters.
Gradient Ascent (GA): In contrast to the gradient descent during the pre-training phase, we maximize the negative log-likelihood loss on the forget corpus. This approach aims to steer the model away from its initial predictions, facilitating the process of unlearning.
Direct Preference Optimization (DPO): We apply preference optimization to enable the model to generate incorrect target knowledge. DPO requires positive and negative examples to train the model. For the positive example, we sample it from the counterfactual corpus, which consists of intentionally fabricated descriptions generated by the model about the target. For the negative example, we sample it from the synthetic forget corpus.
Negative Preference Optimization (NPO): NPO is a simple drop-in fix of the GA loss. Compared to DPO, NPO retains only the negative examples without any positive examples.
Rejection Tuning (RT): First, we have the model generate some questions related to the unlearning targets, then replace its responses with “I do not know the answer.”. Then, we use this refusal data to fine-tune the model so that it can reject questions related to the target.

Forget Corpus Generation

We have provided the forget corpus for both Llama-3-8B-Instruct and Phi-3-mini-4k-instruct to facilitate reproducibility.

from datasets import load_dataset
train_positive_llama3 = load_dataset("jinzhuoran/RWKU", 'train_positive_llama3')['train'] # For GA and NPO
train_pair_llama3 = load_dataset("jinzhuoran/RWKU", 'train_pair_llama3')['train'] # For DPO
train_refusal_llama3 = load_dataset("jinzhuoran/RWKU", 'train_refusal_llama3')['train'] # For RT

train_positive_phi3 = load_dataset("jinzhuoran/RWKU", 'train_positive_phi3')['train'] # For GA and NPO
train_pair_phi3 = load_dataset("jinzhuoran/RWKU", 'train_pair_phi3')['train'] # For DPO
train_refusal_phi3 = load_dataset("jinzhuoran/RWKU", 'train_refusal_phi3')['train'] # For RT

Additionally, you can construct your own forget corpus to explore new methods and models. We have included our generation script for reference. Please feel free to explore better methods for generating forget corpus.

cd generation
python pair_generation.py # For GA, DPO and NPO
python question_generation.py # For RT

Evaluating Models

To evaluate the model original performance before unlearning.

cd LLaMA-Factory/scripts
bash run_original.sh

Unlearning Models

We adapt LLaMA-Factory to train the model. We provide several scripts to run various unlearning methods.

Single-sample Unlearning Setting

To run the In-Context Unlearning (ICU) method on Llama-3-8B-Instruct.

cd LLaMA-Factory
bash scripts/full/run_icu.sh

To run the Gradient Ascent (GA) method on Llama-3-8B-Instruct.

cd LLaMA-Factory
bash scripts/full/run_ga.sh

To run the Direct Preference Optimization (DPO) method on Llama-3-8B-Instruct.

cd LLaMA-Factory
bash scripts/full/run_dpo.sh

To run the Negative Preference Optimization (NPO) method on Llama-3-8B-Instruct.

cd LLaMA-Factory
bash scripts/full/run_npo.sh

To run the Rejection Tuning (RT) method on Llama-3-8B-Instruct.

cd LLaMA-Factory
bash scripts/full/run_rt.sh

Batch-sample Unlearning Setting

To run the In-Context Unlearning (ICU) method on Llama-3-8B-Instruct.

cd LLaMA-Factory
bash scripts/batch/run_icu.sh

To run the Gradient Ascent (GA) method on Llama-3-8B-Instruct.

cd LLaMA-Factory
bash scripts/batch/run_ga.sh

To run the Direct Preference Optimization (DPO) method on Llama-3-8B-Instruct.

cd LLaMA-Factory
bash scripts/batch/run_dpo.sh

To run the Negative Preference Optimization (NPO) method on Llama-3-8B-Instruct.

cd LLaMA-Factory
bash scripts/batch/run_npo.sh

To run the Rejection Tuning (RT) method on Llama-3-8B-Instruct.

cd LLaMA-Factory
bash scripts/batch/run_rt.sh

LoRA Unlearning Setting

Please set --finetuning_type lora and --lora_target q_proj,v_proj.

Partial-layer Unlearning Setting

Please set --train_layers 0-4.

Experimental Results

Results of main experiment on LLaMA3-Instruct (8B).

Results of main experiment on Phi-3 Mini-4K-Instruct (3.8B).

Citation

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

@misc{jin2024rwku,
    title={RWKU: Benchmarking Real-World Knowledge Unlearning for Large Language Models},
    author={Zhuoran Jin and Pengfei Cao and Chenhao Wang and Zhitao He and Hongbang Yuan and Jiachun Li and Yubo Chen and Kang Liu and Jun Zhao},
    year={2024},
    eprint={2406.10890},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

jinzhuoran / RWKU

readme