The Powerful Multi-modal LLM Family

for OCR-free Document Understanding

Alibaba Group

📢 News

• 🔥🔥🔥 [2024.9.28] We have released the training data, inference code and evaluation code of DocOwl2 on both HuggingFace 🤗 and ModelScope .
• 🔥🔥🔥 [2024.9.20] Our paper DocOwl 1.5 and TinyChart is accepted by EMNLP 2024.
• 🔥🔥🔥 [2024.9.06] We release the arxiv paper of mPLUG-DocOwl 2, a SOTA 8B Multimodal LLM on OCR-free Multipage Document Understanding, each document image is encoded with just 324 tokens!
• 🔥🔥 [2024.7.16] Our paper PaperOwl is accepted by ACM MM 2024.
• [2024.5.08] We have released the training code of DocOwl1.5 supported by DeepSpeed. You can now finetune a stronger model based on DocOwl1.5!
• [2024.4.26] We release the arxiv paper of TinyChart, a SOTA 3B Multimodal LLM for Chart Understanding with Program-of-Throught ability (ChartQA: 83.6 > Gemin-Ultra 80.8 > GPT4V 78.5). The demo of TinyChart is available on HuggingFace 🤗. Both codes, models and data are released in TinyChart.
• [2024.4.3] We build demos of DocOwl1.5 on both ModelScope and HuggingFace 🤗, supported by the DocOwl1.5-Omni. The source codes of launching a local demo are also released in DocOwl1.5.
• [2024.3.28] We release the training data (DocStruct4M, DocDownstream-1.0, DocReason25K), codes and models (DocOwl1.5-stage1, DocOwl1.5, DocOwl1.5-Chat, DocOwl1.5-Omni) of mPLUG-DocOwl 1.5 on both HuggingFace 🤗 and ModelScope .
• [2024.3.20] We release the arxiv paper of mPLUG-DocOwl 1.5, a SOTA 8B Multimodal LLM on OCR-free Document Understanding (DocVQA 82.2, InfoVQA 50.7, ChartQA 70.2, TextVQA 68.6).
• [2024.01.13] Our Scientific Diagram Analysis dataset M-Paper has been available on both HuggingFace 🤗 and ModelScope , containing 447k high-resolution diagram images and corresponding paragraph analysis.
• [2023.10.13] Training data, models of mPLUG-DocOwl/UReader has been open-sourced.
• [2023.10.10] Our paper UReader is accepted by EMNLP 2023.

• [2023.07.10] The demo of mPLUG-DocOwl on ModelScope is avaliable.
• [2023.07.07] We release the technical report and evaluation set of mPLUG-DocOwl.

🤖 Models

• mPLUG-DocOwl2 (Arxiv 2024) - mPLUG-DocOwl2: High-resolution Compressing for OCR-free Multi-page Document Understanding

• mPLUG-DocOwl1.5 (EMNLP 2024) - mPLUG-DocOwl 1.5: Unified Structure Learning for OCR-free Document Understanding

• TinyChart (EMNLP 2024) - TinyChart: Efficient Chart Understanding with Visual Token Merging and Program-of-Thoughts Learning

• mPLUG-PaperOwl (ACM MM 2024) - mPLUG-PaperOwl: Scientific Diagram Analysis with the Multimodal Large Language Model

• UReader (EMNLP 2023) - UReader: Universal OCR-free Visually-situated Language Understanding with Multimodal Large Language Model

• mPLUG-DocOwl (Arxiv 2023) - mPLUG-DocOwl: Modularized Multimodal Large Language Model for Document Understanding

📺 Online Demo

Note: The demo of HuggingFace is not as stable as ModelScope because the GPU in ZeroGPU Spaces of HuggingFace is dynamically assigned.

📖 DocOwl 1.5

• 🤗 HuggingFace Space

• ModelScope Space

📈 TinyChart-3B

• 🤗 HuggingFace Space

🌰 Cases

Related Projects

• mPLUG.
• mPLUG-2.
• mPLUG-Owl