Meme Classification on MAMI
0.799AccuracyQwen2-VL-7B
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Qwen2-VL-7BEvaluation Protocol=RA-HMD2025.02 | 0.799 | 0.904 | |
| LLaVA-1.5-7BEvaluation Protocol=RA-HMD2025.02 | 0.797 | 0.912 | |
| GPT-4o2025.02 | 0.794 | — | |
| Qwen2-VL-2BEvaluation Protocol=RA-HMD2025.02 | 0.794 | 0.893 | |
| RGCLTraining=Supervised fine-tuned2025.02 | 0.784 | 0.894 | |
| RA-HMDbackbone=Qwen2-VL-7B, fine-tuning set=HatefulMemes, protocol=RKC2025.02 | 0.756 | 81.4 | |
| LOREHMmode=low-resourced2025.02 | 0.754 | — | |
| HateCLIPperTraining=Supervised fine-tuned2025.02 | 0.748 | 0.872 | |
| RA-HMDbackbone=LLaVA-1.5-7B, fine-tuning set=HatefulMemes, protocol=RKC2025.02 | 0.745 | 80 | |
| LLaVA-1.5-7BEvaluation Protocol=SFT2025.02 | 0.739 | 0.86 | |
| Pro-Cap2025.02 | 0.736 | 0.838 | |
| GPT-4omode=low-resourced2025.02 | 0.729 | — | |
| Qwen2-VL-7BEvaluation Protocol=SFT2025.02 | 0.724 | 0.826 | |
| RA-HMDbackbone=Qwen2-VL-2B, fine-tuning set=HatefulMemes, protocol=RKC2025.02 | 0.711 | 74.8 | |
| Qwen2-VL-2BEvaluation Protocol=SFT2025.02 | 0.686 | 0.777 | |
| CLIPTraining=Supervised fine-tuned2025.02 | 0.684 | 0.777 | |
| Qwen2-VL-7BEvaluation Protocol=4-shot2025.02 | 0.661 | 0.734 | |
| Qwen2-VL-2Bfine-tuning set=HatefulMemes, protocol=SFT + few-shot2025.02 | 0.66 | 73.8 | |
| Qwen2-VL-7Bfine-tuning set=HatefulMemes, protocol=SFT + few-shot2025.02 | 0.66 | 73.4 | |
| LLaVA-1.5-7BEvaluation Protocol=4-shot2025.02 | 0.627 | 0.681 | |
| RGCLfine-tuning set=HatefulMemes2025.02 | 0.624 | 67.8 | |
| Qwen2-VL-7Bfine-tuning set=HatefulMemes, protocol=SFT + zero-shot2025.02 | 0.619 | 71.1 | |
| LLaVA-1.5-7Bfine-tuning set=HatefulMemes, protocol=SFT + zero-shot2025.02 | 0.611 | 69.1 | |
| Mod-Hatemode=low-resourced2025.02 | 0.61 | 67.4 | |
| Qwen2-VL-2BEvaluation Protocol=4-shot2025.02 | 0.588 | 0.648 | |
| Qwen2-VL-7BEvaluation Protocol=zero-shot2025.02 | 0.585 | 0.762 | |
| LLaVA-1.5-7BEvaluation Protocol=zero-shot2025.02 | 0.583 | 0.676 | |
| Qwen2-VL-2Bfine-tuning set=HatefulMemes, protocol=SFT + zero-shot2025.02 | 0.573 | 66.4 | |
| Qwen2-VL-2BEvaluation Protocol=zero-shot2025.02 | 0.51 | 0.672 | |
| LLaVA-1.5-7Bfine-tuning set=HatefulMemes, protocol=SFT + few-shot2025.02 | 0.501 | 65.5 |