tyodd/SafeGuard-VL-RL
tyodd/SafeGuard-VL-RL is a 7 billion parameter visual safety guardrail model built on Qwen2.5-VL-7B-Instruct, designed for policy-aware unsafe-image detection. It is trained using Group Relative Policy Optimization (GRPO) on the LlavaGuard dataset, enabling it to interpret natural language safety policies and make context-sensitive judgments. This model excels at cross-policy generalization, allowing it to adapt to unseen safety policies with varying category definitions, while maintaining general Visual Question Answering (VQA) capabilities.
Loading preview...
SafeGuard-VL-RL: Policy-Aware Visual Safety Guardrail
SafeGuard-VL-RL is a 7 billion parameter visual safety guardrail model developed by tyodd, based on the Qwen2.5-VL-7B-Instruct architecture. Its core innovation lies in its ability to perform policy-aware unsafe-image detection, making it distinct from traditional static classifiers. The model was trained using Group Relative Policy Optimization (GRPO) on the AIML-TUDA/LlavaGuard dataset, focusing on reasoning about image compliance or violation against a given policy.
Key Capabilities
- Policy-aware safety judgment: Interprets natural language safety policies to make context-sensitive decisions on image safety, rather than relying on fixed categories.
- Robust instruction following: Integrates safety guardrail functionality while preserving general Visual Question Answering (VQA) capabilities.
- Cross-policy generalization: Demonstrates the ability to generalize to new, unseen safety policies with different category definitions, enhancing adaptability.
Good for
- Research in visual safety and policy-adaptive AI: Specifically designed for studies on policy-adaptive image guardrails, as detailed in the CVPR 2026 paper "Towards Policy-Adaptive Image Guardrail: Benchmark and Method".
- Developing dynamic content moderation systems: Useful for scenarios requiring flexible safety judgments based on evolving or custom safety policies.
- Exploring RL-based safety alignment: Provides a practical example of applying Group Relative Policy Optimization for visual safety alignment.