|
--- |
|
license: apache-2.0 |
|
--- |
|
# Model Card for Model ID |
|
|
|
|
|
## Welcome |
|
If you find this model helpful, please *like* this model and star us on https://github.com/LianjiaTech/BELLE ! |
|
|
|
## 📝Belle-VL |
|
### 背景介绍 |
|
**社区目前已经有很多多模态大语言模型相关开源工作,但大多以英文能力为主,比如[LLava](https://github.com/haotian-liu/LLaVA),[CogVLM](https://github.com/THUDM/CogVLM)等,而中文多模态大语言模型比如[VisualGLM-6B](https://github.com/THUDM/VisualGLM-6B)、[Qwen-VL](https://github.com/QwenLM/Qwen-VL)的语言模型基座均较小,实际应用中很难兼顾视觉和语言能力,因此Belle-VL选择基于更强的语言模型基座来扩展模型的视觉能力,为社区提供更加灵活的选择。** |
|
|
|
### 模型简介 |
|
在模型结构方面,我们主要参考的[Qwen-VL](https://github.com/QwenLM/Qwen-VL)模型,原始Qwen-VL是基于Qwen7B模型训练而来,基座能力相对较弱,因此Belle-VL将语言模型扩展成了[Qwen14B-chat](https://huggingface.co/Qwen/Qwen-14B-Chat),在中文语言能力和视觉能力方面可以兼顾,具备更好的扩展性。 |
|
|
|
### 训练策略 |
|
原始Qwen-vl采用了三阶段的训练方式,包括预训练、多任务训练和指令微调,依赖较大的数据和机器资源。受LLava1.5的启发,多模态指令微调比预训练更加重要,因此我们采用了两阶段的训练方式,如下图所示: |
|
![Traing_stage](./train.png) |
|
|
|
### 训练数据 |
|
* **预训练数据**:预训练数据主要是基于LLava 的[558k](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain)英文指令数据及其对应的中文翻译数据,此外我们还收集了[Flickr30k-CNA](https://zero.so.com/) 以及从[AI Challenger](https://tianchi.aliyun.com/dataset/145781?spm=a2c22.12282016.0.0.5c823721PG2nBW)随机选取的100k数据 |
|
|
|
* **多模态指令数据**:指令微调阶段,数据主要来自[LLava](https://github.com/haotian-liu/LLaVA), [LRV-Instruction](https://github.com/FuxiaoLiu/LRV-Instruction), [LLaVAR](https://github.com/SALT-NLP/LLaVAR),[LVIS-INSTRUCT4V](https://github.com/X2FD/LVIS-INSTRUCT4V)等开源项目,我们也对其中部分数据进行了翻译,在此真诚的感谢他们为开源所做出的贡献! |
|
|
|
### 模型使用 |
|
``` python |
|
import torch |
|
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig |
|
model_dir = '/path/to_finetuned_model/' |
|
img_path = 'you_image_path' |
|
tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True) |
|
model = AutoModelForCausalLM.from_pretrained(model_dir, trust_remote_code=True).eval() |
|
model.generation_config = GenerationConfig.from_pretrained(model_dir, trust_remote_code=True) |
|
question = '详细描述一下这张图' |
|
|
|
query = tokenizer.from_list_format([ |
|
{'image': img_path}, # Either a local path or an url |
|
{'text': question}, |
|
]) |
|
response, history = model.chat(tokenizer, query=query, history=None) |
|
print(response) |
|
|
|
#or |
|
query = f'<img>{img_path}</img>\n{question}' |
|
response, history = model.chat(tokenizer, query=query, history=None) |
|
print(response) |
|
``` |
|
|
|
|
|
### MME Benchmark |
|
[MME](https://github.com/BradyFU/Awesome-Multimodal-Large-Language-Models/tree/Evaluation)是一个针对多模态大型语言模型的全面评估基准。它在总共14个子任务上测量感知和认知能力,包括 |
|
包括存在性、计数、位置、颜色、海报、名人、场景、地标、艺术作品、OCR、常识推理、数值计算、文本翻译和代码推理等。目前最新的BELLE-VL模型在感知评测维度共获得**1620.10**分,超过LLava和Qwen-VL.详情如下: |
|
|
|
| Category | Score | |
|
|------------------------|-------| |
|
| **Perception** | **1620.10** | |
|
| --Existence | 195.00 | |
|
| --Count | 173.33 | |
|
| --Position | 1310.00 | |
|
| --Color | 185.00 | |
|
| --Posters | 160.88| |
|
| --Celebrity | 135.88| |
|
| --Scene | 150.00| |
|
| --Landmark | 169.25 | |
|
| --Artwork | 143.50 | |
|
| --OCR | 177.50 | |
|
|
|
| Category | Score | |
|
|------------------------|-------| |
|
| **Cognition** | **305.36** | |
|
| --Commonsense Reasoning | 132.86| |
|
| --Numerical Calculation | 42.50 | |
|
| --Text Translation | 72.50 | |
|
| --Code Reasoning | 57.00 | |
|
|
|
|
|
## Citation |
|
|
|
Please cite our paper and github when using our code, data or model. |
|
|
|
``` |
|
@misc{BELLE, |
|
author = {BELLEGroup}, |
|
title = {BELLE: Be Everyone's Large Language model Engine}, |
|
year = {2023}, |
|
publisher = {GitHub}, |
|
journal = {GitHub repository}, |
|
howpublished = {\url{https://github.com/LianjiaTech/BELLE}}, |
|
} |
|
``` |