Skip to content

Dataset · ys-zong

Safety alignment dataset for vision-language models

ys-zong/VLGuard

English image and metadata dataset from an ICML 2024 paper for safety fine-tuning of VLMs.

What's inside

Sourced from an ICML 2024 paper, this English dataset contains 1,000 to 10,000 examples designed for safety tuning of large vision-language models. The data is split into training and test sets, with images packaged in train.zip and test.zip and corresponding metadata stored in train.json and test.json.

What you can do with it

  • SFT fine-tuning of vision-language models to defend against unsafe image-based prompts
  • Testing the robustness of multimodal neural networks against harmful visual scenarios
Tasks
Image → text
Languages
English
Size
1K–10K rows
Downloads this month
1.5K
Likes
22
Access
on request on Hugging Face

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("ys-zong/VLGuard")