Dataset · JailbreakV-28K
JailBreakV-28K: MLLM jailbreak benchmark
JailbreakV-28K/JailBreakV-28k
A collection of images and texts for testing multimodal LLM robustness against jailbreak attacks.
What's inside
This is a benchmark for evaluating the robustness of multimodal large language models to jailbreak attacks. The card specifies images and text, support for attack customization in version v0.2; the project was accepted at COLM 2024. Metadata size ranges from 10k to 100k examples, language not specified in metadata.
What you can do with it
- Evaluating multimodal LLM safety
- Red teaming and defense fine-tuning
- Comparing attack prompts
- Tasks
- Text generation, Question answering
- Size
- 10K–100K rows
- Downloads this month
- 12.9K
- Likes
- 92
Models trained on it
- Horizon-Labs/prompt-injection-guard-baseText classification 7.3K
- Horizon-Labs/prompt-injection-guard-smallText classification 5.4K
- codeintegrity-ai/promptguardText classification 1.4K
- mradermacher/Spellguard-1.0-8B-GGUF 754
- Horizon-Labs/prompt-injection-guard-largeText classification 578
- katanemo/Arch-GuardText classification 396
- mradermacher/GPT-Distill-Qwen3-8B-Spellguard-SLERP-GGUF 337
- katanemo/Arch-Guard-cpuText classification 105
How to load it
from datasets import load_dataset
ds = load_dataset("JailbreakV-28K/JailBreakV-28k")