README.md

metadata

task_categories:
  - image-text-to-text
license: mit

Introduction

This repository contains the data for Open-Qwen2VL: Compute-Efficient Pre-Training of Fully-Open Multimodal LLMs on Academic Resources.

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
• 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

@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}
  }