Kai-Fu Lee participated in Zero One Wish, which released a world-class open source multi-modal large model.

Leading the two authoritative lists in Chinese and English, Kai-Fu Zero handed over the multi-modal large model answer sheet!

It is less than three months since the release of its first open source large models Yi-34B and Yi-6B.

The model is called Yi Vision Language (Yi-VL), and it is now officially open source to the world.

belong to the Yi series and also have two versions:

Yi-VL-34B and Yi-VL-6B.

Let’s take a look at two examples first to experience Yi-VL’s performance in diverse scenarios such as graphic and text dialogues:

Yi-VL Each picture was analyzed in detail, not only explaining the content on the sign, but even taking care of the "ceiling".

In Chinese, Yi-VL can also express clearly and methodically accurately:

In addition, the official test results were also given.

Yi-VL-34B has an accuracy of 41.6% on the English data set MMMU, second only to GPT-4V with an accuracy of 55.7%, surpassing a series of multi-modal large models.

On the Chinese data set CMMMU, the accuracy of Yi-VL-34B is 36.5%, which is ahead of the current cutting-edge open source multi-modal models.

#What does Yi-VL look like?

Yi-VL is developed based on the Yi language model. You can see the powerful text understanding capabilities based on the Yi language model. You only need to align the pictures to get a good multi-modal visual language model - this is also One of the core highlights of the Yi-VL model.

In terms of architecture design, the Yi-VL model is based on the open source LLaVA architecture and contains three main modules:

• Vision Transformer (ViT for short) For image encoding, the open source OpenClip ViT-H/14 model is used to initialize the trainable parameters, and by learning to extract features from large-scale "image-text" pairs, the model has the ability to process and understand images.
• The Projection module brings the ability to spatially align image features and text features to the model. This module consists of a multilayer perceptron (Multilayer Perceptron, referred to as MLP) that contains layer normalizations. This design allows the model to more effectively fuse and process visual and text information, improving the accuracy of multi-modal understanding and generation.
The introduction of Yi-34B-Chat and Yi-6B-Chat large language models provides Yi-VL with powerful language understanding and generation capabilities. This part of the model uses advanced natural language processing technology to help Yi-VL deeply understand complex language structures and generate coherent and relevant text output.

△Caption: Yi-VL model architecture design and training method process overview

On

training method, Yi -The training process of the VL model is divided into three stages, aiming to comprehensively improve the model's visual and language processing capabilities.

In the first stage, the ViT and Projection modules are trained using 100 million "image-text" paired data sets.

At this stage, the image resolution is set to 224x224 to enhance ViT’s knowledge acquisition capabilities in specific architectures while achieving efficient alignment with large language models.

In the second stage, the image resolution of ViT is increased to 448x448, making the model better at recognizing complex visual details. About 25 million "image-text" pairs are used in this stage.

In the third stage, the parameters of the entire model are opened for training, with the goal of improving the model's performance in multi-modal chat interaction. The training data covers diverse data sources, with a total of approximately 1 million "image-text" pairs, ensuring the breadth and balance of the data.

The zero-yiwu technical team also verified that it can quickly train efficient images based on the Yi language model's powerful language understanding and generation capabilities using other multi-modal training methods such as BLIP, Flamingo, EVA, etc. A multimodal graphic-text model for understanding and smoothing graphic-text dialogue.

Yi series models can be used as base language models for multi-modal models, providing a new option for the open source community. At the same time, the zero-one-things multi-modal team is exploring multi-modal pre-training from scratch to approach and surpass GPT-4V faster and reach the world's first echelon level.

Currently, the Yi-VL model has been opened to the public on platforms such as Hugging Face and ModelScope. Users can personally experience the performance of this model in diverse scenarios such as graphic and text dialogues.

Beyond a series of large multi-modal models

In the new multi-modal benchmark test MMMU, both versions Yi-VL-34B and Yi-VL-6B performed well.

MMMU (full name Massive Multi-discipline Multi-modal Understanding & Reasoning Massive Multi-discipline Multi-modal Understanding and Reasoning) The data set contains 11,500 subjects from six core disciplines(Art & Design, Business, Science, Health & Medicine, Humanities & Social Sciences, and Technology & Engineering) questions involving highly heterogeneous image types and intertwined textual image information pose challenges to the model's advanced perception and reasoning capabilities met extremely high demands.

Yi-VL-34B successfully surpassed a series of multi-modal large models with an accuracy of 41.6% on this test set, second only to GPT-4V (55.7%), showing strong ability to understand and apply interdisciplinary knowledge.

Similarly, on the CMMMU data set created for the Chinese scene, the Yi-VL model shows the unique advantage of "understanding Chinese people better".

CMMMU contains about 12,000 Chinese multi-modal questions derived from university exams, tests and textbooks.

Among them, GPT-4V has an accuracy of 43.7% on this test set, followed by Yi-VL-34B with an accuracy of 36.5%, leading the The current cutting-edge open source multimodal model.

Project address:
[1]https://huggingface.co/01-ai
[2] https://www.modelscope.cn/organization/01ai

The above is the detailed content of Kai-Fu Lee participated in Zero One Wish, which released a world-class open source multi-modal large model.. For more information, please follow other related articles on the PHP Chinese website!