mPLUG-Owl: Modularization Empowers Large Language Models with Multimodality
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2023
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866910389361442816 |
|---|---|
| author | Ye, Qinghao Xu, Haiyang Xu, Guohai Ye, Jiabo Yan, Ming Zhou, Yiyang Wang, Junyang Hu, Anwen Shi, Pengcheng Shi, Yaya Li, Chenliang Xu, Yuanhong Chen, Hehong Tian, Junfeng Qian, Qi Zhang, Ji Huang, Fei Zhou, Jingren |
| author_facet | Ye, Qinghao Xu, Haiyang Xu, Guohai Ye, Jiabo Yan, Ming Zhou, Yiyang Wang, Junyang Hu, Anwen Shi, Pengcheng Shi, Yaya Li, Chenliang Xu, Yuanhong Chen, Hehong Tian, Junfeng Qian, Qi Zhang, Ji Huang, Fei Zhou, Jingren |
| contents | Large language models (LLMs) have demonstrated impressive zero-shot abilities on a variety of open-ended tasks, while recent research has also explored the use of LLMs for multi-modal generation. In this study, we introduce mPLUG-Owl, a novel training paradigm that equips LLMs with multi-modal abilities through modularized learning of foundation LLM, a visual knowledge module, and a visual abstractor module. This approach can support multiple modalities and facilitate diverse unimodal and multimodal abilities through modality collaboration. The training paradigm of mPLUG-Owl involves a two-stage method for aligning image and text, which learns visual knowledge with the assistance of LLM while maintaining and even improving the generation abilities of LLM. In the first stage, the visual knowledge module and abstractor module are trained with a frozen LLM module to align the image and text. In the second stage, language-only and multi-modal supervised datasets are used to jointly fine-tune a low-rank adaption (LoRA) module on LLM and the abstractor module by freezing the visual knowledge module. We carefully build a visually-related instruction evaluation set OwlEval. Experimental results show that our model outperforms existing multi-modal models, demonstrating mPLUG-Owl's impressive instruction and visual understanding ability, multi-turn conversation ability, and knowledge reasoning ability. Besides, we observe some unexpected and exciting abilities such as multi-image correlation and scene text understanding, which makes it possible to leverage it for harder real scenarios, such as vision-only document comprehension. Our code, pre-trained model, instruction-tuned models, and evaluation set are available at https://github.com/X-PLUG/mPLUG-Owl. The online demo is available at https://www.modelscope.cn/studios/damo/mPLUG-Owl. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2304_14178 |
| institution | arXiv |
| publishDate | 2023 |
| record_format | arxiv |
| spellingShingle | mPLUG-Owl: Modularization Empowers Large Language Models with Multimodality Ye, Qinghao Xu, Haiyang Xu, Guohai Ye, Jiabo Yan, Ming Zhou, Yiyang Wang, Junyang Hu, Anwen Shi, Pengcheng Shi, Yaya Li, Chenliang Xu, Yuanhong Chen, Hehong Tian, Junfeng Qian, Qi Zhang, Ji Huang, Fei Zhou, Jingren Computation and Language Computer Vision and Pattern Recognition Machine Learning Large language models (LLMs) have demonstrated impressive zero-shot abilities on a variety of open-ended tasks, while recent research has also explored the use of LLMs for multi-modal generation. In this study, we introduce mPLUG-Owl, a novel training paradigm that equips LLMs with multi-modal abilities through modularized learning of foundation LLM, a visual knowledge module, and a visual abstractor module. This approach can support multiple modalities and facilitate diverse unimodal and multimodal abilities through modality collaboration. The training paradigm of mPLUG-Owl involves a two-stage method for aligning image and text, which learns visual knowledge with the assistance of LLM while maintaining and even improving the generation abilities of LLM. In the first stage, the visual knowledge module and abstractor module are trained with a frozen LLM module to align the image and text. In the second stage, language-only and multi-modal supervised datasets are used to jointly fine-tune a low-rank adaption (LoRA) module on LLM and the abstractor module by freezing the visual knowledge module. We carefully build a visually-related instruction evaluation set OwlEval. Experimental results show that our model outperforms existing multi-modal models, demonstrating mPLUG-Owl's impressive instruction and visual understanding ability, multi-turn conversation ability, and knowledge reasoning ability. Besides, we observe some unexpected and exciting abilities such as multi-image correlation and scene text understanding, which makes it possible to leverage it for harder real scenarios, such as vision-only document comprehension. Our code, pre-trained model, instruction-tuned models, and evaluation set are available at https://github.com/X-PLUG/mPLUG-Owl. The online demo is available at https://www.modelscope.cn/studios/damo/mPLUG-Owl. |
| title | mPLUG-Owl: Modularization Empowers Large Language Models with Multimodality |
| topic | Computation and Language Computer Vision and Pattern Recognition Machine Learning |
| url | https://arxiv.org/abs/2304.14178 |