How Many Unicorns Are In This Image? A Safety Evaluation Benchmark For Vision LLMs (Dataset)

Paper: https://arxiv.org/abs/2311.16101

Code: https://github.com/UCSC-VLAA/vllm-safety-benchmark

The full dataset should looks like this:

.
├── ./safety_evaluation_benchmark_datasets//                    
    ├── gpt4v_challenging_set # Contains the challenging test data for GPT4V
        ├── attack_images
        ├── sketchy_images
        ├── oodcv_images
        ├── misleading-attack.json
        ├── sketchy-vqa-challenging.json
        └── oodcv-vqa-counterfactual.json
    ├── redteaming-mislead # Contains the test data for redteaming tasks
        ├── redteaming_attack
            ├── gaussian_noise
            ├── mixattack_eps32
            ├── mixattack_eps64
            ├── sinattack_eps64_dog
            ├── sinattack_eps64_coconut
            ├── sinattack_eps64_spaceship
            └── annotation.json
        └── jailbreak_llm # adversarial suffixes for jailbreaking VLLM through LLM
    └── ood # Contains the test data for OOD scenarios
        ├── sketchy-vqa
            ├── sketchy-vqa.json
            ├── sketchy-challenging.json
        └── oodcv-vqa
            ├── oodcv-vqa.json
            └── oodcv-counterfactual.json