README.md

---
license: apache-2.0
language:
- en
base_model:
- Qwen/Qwen2.5-VL-7B-Instruct
pipeline_tag: image-text-to-text
library_name: transformers
---


<div align="center">

<h1> MoRE: Mixture-of-Retrieval Experts for Reasoning-Guided Multimodal Knowledge Exploitation </h1>


<h5 align="center"> 

<a href='https://arxiv.org/abs/2505.22095'><img src='https://img.shields.io/badge/Paper-Arxiv-red'></a>
<a href='https://huggingface.co/hmhm1229/R1-Router'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-blue'>
<a href='https://huggingface.co/hmhm1229/R1-Router-3B'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-blue'>

Chunyi Peng<sup>1,3</sup>,
Zhipeng Xu<sup>1</sup>,
Zhenghao Liu<sup>1</sup>,
Yishan Li<sup>3</sup>,
Yukun Yan<sup>2</sup>,
Yu Gu<sup>1</sup>
Minghe Yu<sup>1</sup>
Ge Yu<sup>1</sup>
Maosong Sun<sup>2</sup>

<sup>1</sup>Northeastern University, <sup>2</sup>Tsinghua University, <sup>3</sup>ModleBest Inc.

<h5 align="center"> If you find this project useful, please give us a star🌟.
</h5>
</div>


## News
26.04.28 Our Work is now accepted by SIGIR 2026🎉!

25.08.22 We upload [MoRE-3B](https://huggingface.co/hmhm1229/MoRE-3B).

## Environment
For training, answer generation, and evaluation processes:
```bash
conda create -n router python=3.11
conda activate router
pip install requirements_router.txt
```
For retriever and corpus construction processes:
```bash
conda create -n retriever python=3.11
conda activate retriever
pip install requirements_retriever.txt
```

## Corpora Construction
For the text corpus, you can download `enwiki-20241020` from [Huggingface](https://huggingface.co/datasets/hmhm1229/enwiki-20241020). Then preprocess, and index it with the following commands:
```bash
7z x enwiki-20241020-pages-articles-multistream.xml.zip.001 
conda activate retriever
wikiextractor enwiki-20241020-pages-articles-multistream.xml.bz2 -o wiki_extracted
python wiki_preprocess.py
```
For the image corpus, you can directly download [M-BEIR](https://huggingface.co/datasets/TIGER-Lab/M-BEIR). To embed and index it, you can follow the [repository](https://github.com/TIGER-AI-Lab/UniIR)

For the table corpus, you can download, embed and index Open-WikiTable following the [repository](https://github.com/sean0042/Open_WikiTable), or you can download directly the one we have already preprocessed from [here](https://huggingface.co/hmhm1229/table-retriever). 

## Retrievers Preparation
For the Text-Image Retriever, you can directly download [UniIR](https://huggingface.co/TIGER-Lab/UniIR)

For the Table Retriever, you can train it with the help of [repository](https://github.com/sean0042/Open_WikiTable), or you can download it directly from [here](https://huggingface.co/hmhm1229/table-retriever). 

## Datasets
We have prepared all the text datasets in `./datasets`, for images you need to download them from:
- `InfoSeek:` InfoSeek images can be downloaded from [OVEN](https://github.com/open-vision-language/oven/tree/main/image_downloads)
- `Dyn-VQA:` Dynamic VQA images can be downloaded from [DynVQA_en.202412](https://github.com/Alibaba-NLP/OmniSearch/blob/main/dataset/DynVQA_en/DynVQA_en.202412.jsonl)
- `WebQA:` WebQA images can be downloaded from [Google Drive](https://drive.google.com/drive/folders/19ApkbD5w0I5sV1IeQ9EofJRyAjKnA7tb)

## Training
If you do not want to train the model, you can download [R1-Router](https://huggingface.co/hmhm1229/R1-Router) and skip this section to [Evaluation](#evaluation)
### Data Synthesis
If you want to use the ready-to-use synthetic data directly, you can skip this section to [Step-GRPO Training](#step-grpo-training)

First, we need to synthesis the data step by step:
```bash
bash src/data_synthesis/data_synthesis.sh
```
### Step-GRPO Training
Our training framework is based on [EasyR1](https://github.com/hiyouga/EasyR1), only you need to do is to download it and replace some files with the files in `./Easy-R1`.
Then start training with the command:
```bash
conda activate router
bash examples/run_qwen2_5_vl_7b_stepgrpo.sh
```
## Evaluation
We provide the evaluation pipeline for the R1-Router:
```bash
bash evaluation.sh
```
or, you can just evaluate the results we have provided by:
```bash
conda activate router
cd src
python evaluate.py --dataset_name all --method "r1-router3"
```

## Acknowledgement 
Our work is built on the following codebases, and we are deeply grateful for their contributions.
- [EasyR1](https://github.com/hiyouga/EasyR1)
- [UniIR](https://huggingface.co/TIGER-Lab/UniIR)
- [Open-WikiTable](https://github.com/sean0042/Open_WikiTable)
- [OmniSearch](https://github.com/Alibaba-NLP/OmniSearch)

## Citation
We appreciate your citations if you find our paper relevant and useful to your research!
```
@article{peng2025r1,
  title={Learning to Route Queries across Knowledge Bases for Step-wise Retrieval-Augmented Reasoning},
  author={Peng, Chunyi and Xu, Zhipeng and Liu, Zhenghao and Li, Yishan and Yan, Yukun and Wang, Shuo and Liu, Zhiyuan and Gu, Yu and Yu, Minghe and Yu, Ge and Sun, Maosong},
  year={2025}
  url={https://arxiv.org/abs/2505.22095}, 
}
```

## Contact Us
If you have questions, suggestions, or bug reports, please email us. We will try our best to help you.
```
hm.cypeng@gmail.com
```