Multimodal Safety on VLGuard
95.11F1 ScoreSingGuard-8B
Evaluation results
| Method | Links | |
|---|---|---|
| SingGuard-8BCategory=Our Models, Evaluation Protocol=Full triple (query, image, response)2026.06 | 95.11 | |
| SingGuard-2BCategory=Our Models, Evaluation Protocol=Full triple (query, image, response)2026.06 | 93.91 | |
| SingGuard-4BCategory=Our Models, Evaluation Protocol=Full triple (query, image, response)2026.06 | 92.96 | |
| GuardReasoner-VL-7BCategory=Open-Source Models, Evaluation Protocol=Full triple (query, image, response)2026.06 | 90.33 | |
| LLaVAShieldCategory=Open-Source Models, Evaluation Protocol=Full triple (query, image, response)2026.06 | 90.11 | |
| GPT-5.1Category=Closed-Source Models, Evaluation Protocol=Full triple (query, image, response)2026.06 | 89.4 | |
| Qwen3-VL-235BCategory=Open-Source Models, Evaluation Protocol=Full triple (query, image, response)2026.06 | 88.35 | |
| Gemini3-ProCategory=Closed-Source Models, Evaluation Protocol=Full triple (query, image, response)2026.06 | 85.7 | |
| ShieldGemma-2Category=Open-Source Models, Evaluation Protocol=Full triple (query, image, response)2026.06 | 69.67 | |
| LlamaGuard4Category=Open-Source Models, Evaluation Protocol=Full triple (query, image, response)2026.06 | 65.14 | |
| Qwen3-VL-4BCategory=Open-Source Models, Evaluation Protocol=Full triple (query, image, response)2026.06 | 57.02 | |
| Qwen3-VL-8BCategory=Open-Source Models, Evaluation Protocol=Full triple (query, image, response)2026.06 | 55.35 | |
| SafeGuard-VLCategory=Open-Source Models, Evaluation Protocol=Full triple (query, image, response)2026.06 | 52.24 | |
| LlavaGuardCategory=Open-Source Models, Evaluation Protocol=Full triple (query, image, response)2026.06 | 50.03 | |
| LlamaGuard3-Vision-11BCategory=Open-Source Models, Evaluation Protocol=Full triple (query, image, response)2026.06 | 34.16 |