You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

ViSafe-Eval generated images: Terms of Use

  • This dataset is released for research on the safety of vision-language models, such as evaluating and improving how they handle harmful requests.
  • Do not use it to bypass or weaken the safety measures of AI systems, or to produce or spread harmful content.
  • The images show the text of harmful requests, and some depict violent or sexual content. Please handle them with care.
  • Each image follows the license of the intent it was made from: CC BY-NC 4.0 for the MM-SafetyBench intents (non-commercial) and MIT for the HarmBench intents. Please follow these licenses and cite our paper when you use the dataset.
  • The dataset is provided as is. The creators are not responsible for any damage caused by its use, and may change or withdraw the dataset without notice.

Log in or Sign Up to review the conditions and access this dataset content.

[EMNLP 2026] Do VLMs Share Safety Neurons Across Modalities?

Jiaxuan Li*, Jiahao Zhang*, Duc Minh Vo, Huy H. Nguyen, Pride Kavumba, Koki Wataoka
SB Intuitions Corp., Tokyo, Japan
* Equal contribution

      

⚠️ Warning: the images show the text of harmful requests, and some depict violent or sexual content. They are intended for safety research only.

πŸ‘€ Overview

This repository holds the images we made for ViSafe-Eval, the evaluation benchmark of our EMNLP 2026 paper Do VLMs Share Safety Neurons Across Modalities?.

Vision-language models (VLMs) can comply with harmful requests that arrive through images, even when their LLM backbone would refuse the same content in text. ViSafe-Eval tests this gap with 472 harmful intents, 8 image conditions and 3 query types.

Only our own images are in this repository. Images from other datasets are not redistributed. The fetch script in the code repository downloads them from their original sources, adds the images of this archive and builds the full benchmark.

πŸ“‚ Contents

visafe_eval_generated_images.zip holds 1,518 images. The table shows where the images of each condition come from.

Image condition Images in ViSafe-Eval In this archive From the fetch script
natural_harmful 472 102 (FLUX.1-schnell, for the HarmBench intents) 370 (MM-SafetyBench SD images of the same intents)
rendered 472 472 (the intent text rendered as an image)
rendered_handwrite 472 472 (the same in a handwriting style)
rendered_poster 472 472 (the same in a poster style)
safe 472 472 (Places365 validation 238, ImageNet-1k validation 234)
blank, noise 1 each made by the script
text_only no image
Total 1,518

Each file is named after its intent id in data/visafe_eval/intents.jsonl of the code repository:

visafe_eval_generated_images.zip
β”œβ”€β”€ LICENSE_AND_README.txt
└── images/
    β”œβ”€β”€ rendered/              # visafe_0000.jpg, ... (472)
    β”œβ”€β”€ rendered_handwrite/    # (472)
    β”œβ”€β”€ rendered_poster/       # (472)
    └── natural_harmful/       # FLUX.1-schnell images of the HarmBench intents (102)

πŸ“¦ Usage

Access requires accepting the terms of use on this page. Then log in with hf auth login (or set HF_TOKEN) and run the following from the root of the code repository:

python data/build/fetch_visafe_eval.py
# -> data/visafe_eval/images/ (the 7 conditions with an image) and data/visafe_eval/metadata.jsonl

The script downloads this archive and checks every image against data/visafe_eval/generated_images.sha256. To use a local copy of the archive, pass it with --generated-zip <path>. metadata.jsonl has one row per sample with its query, image path, intent, category and condition. Some source datasets are gated and need an approved Hugging Face token. The README of the code repository lists them.

πŸ“œ License

Each image is made from the text of one intent and follows the license of that intent's source: CC BY-NC 4.0 for the MM-SafetyBench intents (1,110 images) and MIT for the HarmBench intents (408 images). FLUX.1-schnell is released under Apache-2.0.

πŸ“§ Contact

If you have any questions, please contact jiaxuan.li@sbintuitions.co.jp.

πŸ“Œ BibTeX & Citation

If you find this dataset useful, please consider citing our work:

@inproceedings{li-etal-2026-vlms,
    title     = "Do {VLM}s Share Safety Neurons Across Modalities?",
    author    = "Li, Jiaxuan and Zhang, Jiahao and Vo, Duc Minh and Nguyen, Huy H. and Kavumba, Pride and Wataoka, Koki",
    booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
    year      = "2026",
    publisher = "Association for Computational Linguistics",
    url       = "https://aclanthology.org/2026.emnlp-main.471/",
    pages     = "10822--10837",
}
Downloads last month
26

Paper for sbintuitions/ViSafe-Eval-generated