PolicyShiftGuard/PolicyShiftGuard-3B
PolicyShiftGuard-3B is a 3 billion parameter policy-conditioned image guardrail model based on Qwen2.5-VL-3B, designed to assess if an image violates a given policy bundle. It returns a structured safe/unsafe decision along with the violated risk category. This model is specifically trained for adaptive image moderation and robustness under policy shifts, making it suitable for research in policy-conditioned multimodal safety.
Loading preview...
PolicyShiftGuard-3B Overview
PolicyShiftGuard-3B is a 3-billion parameter image guardrail model built upon the Qwen2.5-VL-3B architecture. Its core function is to evaluate images against a provided set of policies, determining if any violations occur. The model outputs a structured decision indicating whether an image is safe or unsafe, and if unsafe, specifies the two-digit risk category and a brief reason.
Key Capabilities
- Policy-Conditioned Image Moderation: Assesses image safety based on explicit, supplied policy bundles rather than fixed rules.
- Structured Output: Provides clear
true | <risk category> | <reason>orfalse | <reason>responses. - Adaptability: Trained with randomized policy SFT and boundary-pair policy adaptation to handle policy shifts.
Good for
- Research in Multimodal Safety: Ideal for studies on policy-conditioned image safety and adaptive moderation systems.
- Evaluating Robustness: Useful for investigating model performance under changing policy definitions.
- Adaptive Image Guardrails: Developing systems that can dynamically adjust moderation criteria based on specific policy requirements.
This model is a research checkpoint and should be evaluated with explicit policy bundles. It may have limitations with policies, languages, visual domains, or deployment settings not covered in its training benchmark.