AIML-TUDA/QwenGuard-v1.2-3B

QwenGuard-v1.2-3B: A Vision Safeguard Model

QwenGuard-v1.2-3B is a specialized 3 billion parameter vision safeguard model developed by AIML-TUDA. It is built upon the Qwen/Qwen2.5-VL-3B-Instruct architecture and is trained on the LlavaGuard-DS dataset. This model's primary function is to evaluate images against a predefined safety policy, delivering a safety rating, categorizing potential violations, and providing a comprehensive rationale for its assessment.

Key Capabilities

• Image Safety Assessment: Determines if an image complies with a given safety policy.
• Policy-Based Categorization: Assigns a specific safety category (e.g., Hate, Violence, Sexual Content, Nudity, Criminal Planning, Weapons/Substance Abuse, Self-Harm, Animal Cruelty, Disasters/Emergencies) or 'None applying'.
• Detailed Rationale Generation: Provides a clear, reasoned explanation for its safety rating, referencing specific policy guidelines.
• Improved Reasoning: Features enhanced reasoning capabilities within the generated rationales compared to its base model.
• Customizable Policies: Operates based on a detailed, configurable safety taxonomy, allowing users to define specific criteria for assessment.

Good For

• Automated Content Moderation: Ideal for automatically screening visual content for policy violations.
• Enhancing AI Safety: Contributes to building safer AI systems by providing a robust mechanism for visual content governance.
• Research in Vision Safeguarding: Serves as an open-source framework for further research and development in VLM-based safety.
• Compliance and Risk Management: Assists organizations in maintaining compliance with content safety standards and mitigating risks associated with harmful visual content.