Harm Recognition on COCO Crime Scene slice
93.9F1 ScoreSingGuard-4B
Evaluation Results
| Method | Links | |
|---|---|---|
| SingGuard-4BSetting=Pixel-only2026.06 | 93.9 | |
| SingGuard-8BSetting=Pixel-only2026.06 | 92.31 | |
| SingGuard-2BSetting=Pixel-only2026.06 | 92.25 | |
| GPT-5.1Setting=Pixel-only2026.06 | 83.97 | |
| LLaVAShieldSetting=Pixel-only2026.06 | 80.95 | |
| GuardReasoner-VLSetting=Pixel-only2026.06 | 80.93 | |
| SafeGuard-VL-RLSetting=Pixel-only2026.06 | 74.44 | |
| Gemini3-ProSetting=Pixel-only2026.06 | 73.61 | |
| LlamaGuard3-VisionSetting=Pixel-only2026.06 | 71.88 | |
| LlavaGuardSetting=Pixel-only2026.06 | 67.13 | |
| ShieldGemma-2Setting=Pixel-only2026.06 | 66.67 | |
| Qwen3-VL-235BSetting=Pixel-only2026.06 | 62.03 | |
| LlamaGuard4Setting=Pixel-only2026.06 | 61.5 | |
| Qwen3-VL-8BSetting=Pixel-only2026.06 | 54.31 | |
| Qwen3-VL-4BSetting=Pixel-only2026.06 | 51.73 |