Offense Classification on MMSD Offense style 2.0 (test)
61.49AccuracySFT on Combined → GRPO (Combined)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFT on Combined → GRPO (Combined)Prompting=zero-shot, Setup=Combined2026.01 | 61.49 | 0.79 | |
| Gemini 2.5 FlashPrompting=zero-shot2026.01 | 60.95 | 0.77 | |
| SFT on Combined → GRPO (Style specific)Prompting=zero-shot, Setup=Style specific2026.01 | 60.27 | 0.77 | |
| LLaMA 90B Vision InstructPrompting=zero-shot2026.01 | 55.47 | 0.75 | |
| Qwen2.5-VL 32B InstructPrompting=zero-shot2026.01 | 49.37 | 0.52 | |
| Phi-4 MultimodalPrompting=zero-shot2026.01 | 48.74 | 0.64 | |
| Gemma 3 27BPrompting=zero-shot2026.01 | 45.33 | 0.57 | |
| LLaVA 1.5 7BPrompting=zero-shot2026.01 | 43.22 | 0.52 |