Meta刚开源llama 3.2多模态,就被打败了!
2个系列:
- 多模态:Llama 3.2 11B 和 90B,支持视觉多模态,LLama终于有了眼睛!
- 端侧小模型:LLama 3.2 1B 和 3B
对于新增的多模态模型,只新增了图像编码器,将其集成到预训练模型中,没有更新语言模型参数,即插即用!
对于 Llama 3.2 中的 1B 和 3B 模型,直接使用8B、70B的logits蒸馏,比较常见的蒸馏方式,就是废卡。
图片
虽然 LLama 系列终于有了视觉。但是,今天 allenai 开源了多模态 Molmo 72B 和 7B 模型。并且在视觉方面的表现全面超过了 3.2,太卷了~
下表是目前知名多模态模型的横向对比,可能存在错误(claude生成的~)
Benchmark | Molmo-72B | Molmo-7B-D | Molmo-7B-O | MolmoE-1B | Llama 3.2 11B | Llama 3.2 90B | Qwen-VL-72B | GPT-4o | Claude-3.5 Sonnet | Qwen2-VL-7B | GPT-4o-mini | InternVL2-8B | MiniCPM-V 2.6 |
AI2D | 96.3 | 93.2 | 90.7 | 86.4 | 62.4 | 75.3 | - | - | - | - | - | - | - |
ChartQA | 87.3 | 84.1 | 80.4 | 78.0 | 83.4 | 85.5 | 88.3 | 85.7 | 90.8 | 83.0 | - | 83.3 | - |
VQAv2 | 86.5 | 85.6 | 85.3 | 83.9 | 75.2 | 78.1 | - | - | - | - | - | - | - |
DocVQA | 93.5 | 92.2 | 90.8 | 77.7 | 88.4 | 90.1 | 96.5 | 92.8 | 95.2 | 94.5 | - | 91.6 | 90.8 |
InfoVQA | 81.9 | 72.6 | 70.0 | 53.9 | 43.2 | 56.8 | 84.5 | - | - | 76.5 | - | 74.8 | - |
TextVQA | 83.1 | 81.7 | 80.4 | 78.8 | 73.1 | 73.5 | 85.5 | - | - | 84.3 | - | 77.4 | 80.1 |
RealWorldQA | 75.2 | 70.7 | 67.5 | 60.4 | N/A | N/A | 77.8 | 75.4 | 60.1 | 70.1 | - | 64.4 | - |
MMMU | 54.1 | 45.3 | 39.3 | 34.9 | 41.7 | 49.3 | 64.5 | 69.1 | 68.3 | 54.1 | 60.0 | 51.8 | 49.8 |
MathVista | 58.6 | 51.6 | 44.5 | 34.0 | 51.5 | 57.3 | 70.5 | 63.8 | 67.7 | 58.2 | 52.4 | 58.3 | 60.6 |
OCRBench | - | - | - | - | - | - | 877 | 736 | 788 | 845 | 785 | 794 | 852 |
MTVQA | - | - | - | - | - | - | 30.9 | 27.8 | 25.7 | 26.3 | - | - | - |
VCR_un easy | - | - | - | - | - | - | 91.93 | 91.55 | 63.85 | 89.70 | 83.60 | - | 73.88 |
MMBench-EN | - | - | - | - | - | - | 86.5 | 83.4 | 79.7 | 83.0 | - | 81.7 | - |
MMStar | - | - | - | - | - | - | 68.3 | 63.9 | 62.2 | 60.7 | 54.8 | 61.5 | 57.5 |
HallBench | - | - | - | - | - | - | 58.1 | 55.0 | 49.9 | 50.6 | 46.1 | 45.2 | 48.1 |
Video-MME | - | - | - | - | - | - | 71.2/77.8 | 71.9/71.2 | 75.0/81.3 | 63.3/69.0 | - | 54.0/56.9 | 60.9/63.6 |
本文转载自 NLP前沿,作者: 热爱AI的