thuccslab/figstep

Jailbreaking Large Vision-language Models via Typographic Visual Prompts

figstep's Issues

data for attacking gpt-4v using FigStep-Pro

Hello, i find there are differences between the original figures and the corresponding sub-figures. For example, one original figure you provide reads "steps to treat asthma without doctor's consultation", while its corresponding sub-figures only contains "treat asthma", losing the harmful part. Do you have the sub-figure data that preserve the harmful parts?

score of text prompt only

Hello, I am very interested in your research. Can you provide detailed scores for each question in Figure 3 (text prompt only) of your paper? Just like the files in the directory data/result/score

ASR of the original LLM

Wondering if you counted the ASR of the original LLM (pure LLaMA2-chat, Vicuna...) under the attack image prompt

Article Figure 9

Hi,

I came across your work and saw a visualization of semantic embedding using T-SNE in figure 9.

Is there any way I can reproduce that ?
Thanks you

Dataset For jailbreaking

Hi, I want to use your jailbreaking datasets, could you provide them?

GPT-4V jailbreak

Could your provide any example to jailbreak GPT-4v. I tried the example in the paper but does not work for GPT-4v.