Open-Qwen2VL-Data

## Introduction

This repository contains the data for [Open-Qwen2VL: Compute-Efficient Pre-Training of Fully-Open Multimodal LLMs on Academic Resources](https://huggingface.co/papers/2504.00595).

Project page: https://victorwz.github.io/Open-Qwen2VL

Code: https://github.com/Victorwz/Open-Qwen2VL

## Dataset
- ccs_ebdataset: CC3M-CC12M-SBU filtered by CLIP, we directly download the webdataset based on the [released of curated subset of BLIP-1](https://github.com/salesforce/BLIP)
- datacomp_medium_dfn_webdataset: DataComp-Medium-128M filtered by DFN, we just select this subset based the uids released by DFN
- datacomp_medium_mlm_filter_su_85_union_dfn_webdataset: DataComp-Medium-128M filtered by DFN union DataComp-Medium-128M filtered by MLM-Filter based on the semantic understanding metric with threshold 85

## Acknowledgement
This work was partially supported by the BioPACIFIC Materials Innovation Platform of the National Science Foundation under Award No. DMR-1933487

## Citation
```bibtex
@article{Open-Qwen2VL,
title={Open-Qwen2VL: Compute-Efficient Pre-Training of Fully-Open Multimodal LLMs on Academic Resources},
author={Wang, Weizhi and Tian, Yu and Yang, Linjie and Wang, Heng and Yan, Xifeng},
journal={arXiv preprint arXiv:2504.00595},
year={2025}
}
```

# 简介
本仓库包含论文《Open-Qwen2VL：基于学术资源的全开源多模态大语言模型（Large Language Model, LLM）的计算高效预训练》的配套数据，论文链接：https://huggingface.co/papers/2504.00595。
项目主页：https://victorwz.github.io/Open-Qwen2VL
代码仓库：https://github.com/Victorwz/Open-Qwen2VL

# 数据集
- ccs_ebdataset：经CLIP筛选后的CC3M-CC12M-SBU数据集，我们直接基于[BLIP-1的精选子集发布内容](https://github.com/salesforce/BLIP)下载对应的webdataset格式数据。
- datacomp_medium_dfn_webdataset：经DFN筛选后的DataComp-Medium-128M数据集，我们仅根据DFN公布的唯一标识符（Unique Identifier, UID）选取该子集。
- datacomp_medium_mlm_filter_su_85_union_dfn_webdataset：将经DFN筛选的DataComp-Medium-128M数据集与基于语义理解指标（阈值设为85）经MLM-Filter筛选的DataComp-Medium-128M数据集取并集后得到的数据集。

# 致谢
本研究部分依托美国国家科学基金会（National Science Foundation）资助的BioPACIFIC材料创新平台（项目编号：DMR-1933487）完成。

# 引用格式
bibtex
@article{Open-Qwen2VL,
title={Open-Qwen2VL: Compute-Efficient Pre-Training of Fully-Open Multimodal LLMs on Academic Resources},
author={Wang, Weizhi and Tian, Yu and Yang, Linjie and Wang, Heng and Yan, Xifeng},
journal={arXiv preprint arXiv:2504.00595},
year={2025}
}