Multimodal AI Risk Benchmark Dataset
Dec 30, 2024
·
1 min read
📸 Gallery
Collaborated with Seoul City University, and industry partners to build a large-scale benchmark dataset for AI safety evaluation.
This dataset covers 35 categories of AI-related risks with 11,480 multimodal instances (text, image, video, audio).
Our lab contributed to the development of risk data for image and video modalities, ensuring high-quality, ethically curated datasets for generative AI safety evaluation.
🏷️ Contribution
- Image Data: 860 instances
- Video Data: 310 instances
Prompt Types included:
- Multiple-Choice
- Q Only
- Multi-Session
- Role-Playing
- Chain-of-Thought
- Expert Prompting
📊 Dataset Overview (Full)
Total Data: 11,480 instances
- Text: 9,560
- Image: 1,160
- Video: 430
- Audio: 330
Risk Categories: 35
Prompt Types: Multiple-Choice, Q Only, Multi-Session, Role-Playing, Chain-of-Thought, Expert Prompting, Rail, Reflection
Authors
Ph.D. AI Researcher | XR Simulation | Explainable AI | Anomaly Detection
I am an AI researcher with a Ph.D. in Computer Science at KAIST, specializing in Generative AI for XR simulations and anomaly detection in safety-critical systems.
My work focuses on Explainable AI (XAI) to enhance transparency and reliability across smart infrastructure, security, and education.
By building multimodal learning approaches and advanced simulation environments, I aim to improve operational safety, immersive training, and scalable content creation.
My work focuses on Explainable AI (XAI) to enhance transparency and reliability across smart infrastructure, security, and education.
By building multimodal learning approaches and advanced simulation environments, I aim to improve operational safety, immersive training, and scalable content creation.