Multimodal Content Moderation on UnsafeBench
76.7AccuracyTool-MCoT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Tool-MCoTFine-tuning Protocol=GRPO2026.03 | 76.7 | 77.3 | |
| Tool-MCoTFine-tuning Protocol=Reasoning LoRA2026.03 | 74.3 | 75.7 | |
| Gemma3-4B-itFine-tuning Protocol=LoRA2026.03 | 70.9 | 74.1 | |
| Gemma3-4B-itFine-tuning Protocol=Base2026.03 | 48.6 | 52.3 |