| license: apache-2.0 | |
| base_model: Qwen/Qwen2.5-VL-3B-Instruct | |
| tags: | |
| - vision-language | |
| - image-safety | |
| - guardrails | |
| - policy-conditioned | |
| - qwen2.5-vl | |
| datasets: | |
| - PolicyShiftBench/PolicyShiftBench | |
| # PolicyShiftGuard-3B | |
| PolicyShiftGuard-3B is a policy-conditioned image guardrail model based on Qwen2.5-VL-3B. It is trained to decide whether an image violates a supplied policy bundle and to return a structured safe/unsafe decision with the violated risk category when applicable. | |
| ## Expected Output Format | |
| ```text | |
| true | <two-digit risk category id> | <short reason> | |
| false | <short reason> | |
| ``` | |
| ## Training Data | |
| This checkpoint is trained with the PolicyShiftBench public data release: | |
| - Dataset: `PolicyShiftBench/PolicyShiftBench` | |
| - Main evaluation splits: ID/adaptive branch and OOD/shift branch | |
| - Training stages: randomized policy SFT followed by boundary-pair policy adaptation | |
| ## Intended Use | |
| Use this model for research on policy-conditioned multimodal safety, adaptive image moderation, and robustness under policy shifts. The model should be evaluated with explicit policy bundles rather than as a fixed universal safety classifier. | |
| ## Limitations | |
| This is a research checkpoint. It may fail under policies, languages, visual domains, or deployment settings not represented in the benchmark. Outputs should not be treated as legal or compliance advice. | |