UCSC-VLAA/vllm-safety-benchmark

[ECCV 2024] Official PyTorch Implementation of "How Many Unicorns Are in This Image? A Safety Evaluation Benchmark for Vision LLMs"

25
/ 100
Experimental

This project helps AI researchers and developers evaluate the safety and robustness of their Vision-Language Models (VLLMs). It takes a VLLM and a variety of challenging image and text datasets as input, then measures how well the VLLM handles out-of-distribution scenarios and red-teaming attacks. The output provides quantifiable metrics on a VLLM's safety performance, helping identify vulnerabilities before deployment.

No commits in the last 6 months.

Use this if you are developing or deploying Vision-Language Models and need to rigorously assess their behavior and safety under challenging and adversarial conditions.

Not ideal if you are looking for a general-purpose VLLM or a tool to generate adversarial attacks without a focus on safety benchmarking.

AI Safety Vision-Language Models Model Evaluation Adversarial Robustness AI Red Teaming
No License Stale 6m No Package No Dependents
Maintenance 0 / 25
Adoption 9 / 25
Maturity 8 / 25
Community 8 / 25

How are scores calculated?

Stars

87

Forks

5

Language

Python

License

Last pushed

Nov 28, 2023

Commits (30d)

0

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/transformers/UCSC-VLAA/vllm-safety-benchmark"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.