Awesome

We present SOLO, a single Transformer architecture for unified vision-language modeling. SOLO accepts both raw image patches (in pixels) and texts as inputs, without using a separate pre-trained vision encoder.

TODO Roadmap

  ✅ Release the instruction tuning data mixture

  ✅ Release the code for instruction tuning

  ✅ Release the pre-training code

  ✅ Release the SOLO model <a href="https://huggingface.co/YangyiYY/SOLO-7B" >🤗 Model (SOLO-7B)</a>

  ✅ Paper on arxiv <a href="https://arxiv.org/abs/2407.06438">📃 Paper</a>

Setup

Clone Repo

git clone https://github.com/Yangyi-Chen/SOLO
git submodule update --init --recursive

Setup Environment for Data Processing

conda env create -f environment.yml
conda activate solo

OR simply

pip install -r requirements.txt

SOLO Inference with Huggingface

Check scripts/notebook/demo.ipynb for an example of performing inference on the model.

Pre-Training

Please refer to PRETRAIN_GUIDE.md for more details about how to perform pre-training. The following table documents the data statistics in pre-training:

Instruction Fine-Tuning

Please refer to SFT_GUIDE.md for more details about how to perform instruction fine-tuning. The following table documents the data statistics in instruction fine-tuning:

Citation

If you use or extend our work, please consider citing our paper.

@article{chen2024single,
  title={A Single Transformer for Scalable Vision-Language Modeling},
  author={Chen, Yangyi and Wang, Xingyao and Peng, Hao and Ji, Heng},
  journal={arXiv preprint arXiv:2407.06438},
  year={2024}
}