MLLMJailbreak-en Dataset

The MLLMJailbreak-en dataset is an English-based multimodal large language model (MLLM) jailbreak attack dataset created from an existing English text dataset using FigStep, MML, and SI-Attack methods.

Contributors

이 데이터셋은 강민석, 김찬영, 이명원, 최희망, 정다흰에 의해 제작되었습니다.

This dataset is made by Minseok Kang, Chanyoung Kim, Myungwon Lee, Heemang Choi, and Dahuin Jung.

Download

import os
import tarfile
from datasets import load_dataset
from huggingface_hub import hf_hub_download
from PIL import Image

repo_id = "HAI-Lab/MLLMJailbreak-en"

cache_dir = "./hf_cache" # For TAR file extraction, using a local cache directory is recommended.
extract_dir = "./images_en"

# 1. download dataset (parquet)
ds = load_dataset(
    repo_id,
    cache_dir=cache_dir,
    # download_mode="force_redownload" # already downloaded
)

# 2. download images_en.tar 
tar_path = hf_hub_download(
    repo_id=repo_id,
    filename="images_en.tar",
    repo_type="dataset",
    cache_dir=cache_dir
)

print(f"[INFO] Downloaded tar: {tar_path}")

# 3. tar extract
os.makedirs(extract_dir, exist_ok=True)

with tarfile.open(tar_path) as tar:
    tar.extractall(extract_dir)

print(f"[INFO] Extracted to: {extract_dir}")

print(ds)

Directory Structure After Installation

working_directory/
├── hf_cache
└── images_en/images_en/
    ├── FigStep/
        └── ...
    ├── MML/
        └── ...
    └── SIAttack/
        └── ...

Dataset Usage

repo_id = "HAI-Lab/MLLMJailbreak-en"
cache_dir = "./hf_cache"

ds = load_dataset(
    repo_id,
    cache_dir=cache_dir,
)

print(ds['train'])

example = ds['train'][0]
print(example)

image_path = example["image"]
print(f"Image path: {image_path}")

img = Image.open(image_path)
img.save("test.png")

Orignal Prompts

The index corresponding to original_prompt_idx and the original prompt can be found in original_prompt.csv under the "Files and versions" section above.

Some indices were excluded during the dataset creation process.

Acknowledgments

This research was supported by the AI Computing Infrastructure Enhancement (GPU Rental Support) User Support Program funded by the Ministry of Science and ICT (MSIT), Republic of Korea (RQT-25-090040).

This project was developed using the following datasets.

walledai/MultiJail

walledai/JailbreakBench

TrustAIRLab/in-the-wild-jailbreak-prompts

The following methods were used in this project.

Figstep by Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang

MML by Yu Wang, Xiaofei Zhou, Yichen Wang, Geyuan Zhang, and Tianxing He

SI-Attack by Shiji Zhao, Ranjie Duan, Fengxiang Wang, Chi Chen, Caixin Kang, Shouwei Ruan, Jialing Tao, YueFeng Chen, Hui Xue, and Xingxing Wei

License

This dataset is released under the MIT License.