AIML-TUDA/QwenGuard-v1.2-7B
AIML-TUDA/QwenGuard-v1.2-7B is a 7 billion parameter vision safeguard model built upon Qwen/Qwen2.5-VL-7B-Instruct, trained on the LlavaGuard-DS dataset. This model assesses images against a provided safety policy, generating a safety rating, category, and a detailed rationale. It specializes in vision-based content moderation, offering improved reasoning capabilities within its rationales for safety assessments. The model is designed for applications requiring automated visual content policy enforcement and safety analysis.
Loading preview...
Model Overview
AIML-TUDA/QwenGuard-v1.2-7B is a 7 billion parameter vision safeguard model developed by AIML-TUDA. It is built on the Qwen/Qwen2.5-VL-7B-Instruct architecture and has been specifically trained on the LlavaGuard-DS dataset.
Key Capabilities
- Vision Safeguarding: Assesses images based on a user-defined safety policy.
- Safety Rating & Categorization: Provides a clear "Safe" or "Unsafe" rating and assigns a specific safety category (e.g., Hate, Violence, Sexual Content, Nudity, Criminal Planning, Weapons, Self-Harm, Animal Cruelty, Disasters).
- Detailed Rationale Generation: Offers improved reasoning capabilities, generating comprehensive justifications for its safety assessments, referencing specific policy guidelines.
- Policy-Driven Analysis: Operates by evaluating visual content against a structured safety taxonomy, allowing for customizable content moderation.
Use Cases
This model is particularly well-suited for:
- Automated content moderation platforms requiring visual safety checks.
- Filtering user-generated image content against predefined safety policies.
- Applications needing explainable AI for visual content safety decisions.
- Research and development in human-centered generative AI and vision safeguarding.
For more technical details, refer to the associated LlavaGuard project website and the LlavaGuard-Arxiv paper.