Introduction to GeoX

GeoX is a multi-modal large model designed for automatic geometric problem solving, utilizing three progressive training stages to enhance diagram understanding and reasoning. In this paper, we validate that the formal vision-language training paradigm is a simple-yet-effective solution for complex mathematical diagram learning.

Data Preparation for GeoX

Step 1. Data for Unimodal Pre-training

You can download our collected diagram images from this link.

Additionally, we use existing geometric text to build a corpus, which is detailed in our paper.

Step 2. Data for Geometry-Language Alignment

To train the GS-Former, please prepare the unified formal annotations and paired images.

Step 3. Data for End-to-End Visual Instruction Tuning

We use the GeoQA, UniGeo, Geometry3K, and PGPS9K datasets for fine-tuning and evaluation:

1. GeoQA: Follow the instructions here to download the GeoQA dataset.
2. UniGeo: Follow the instructions here to download the UniGeo dataset.
3. Geometry3K and PGPS9K: Follow the instructions here to download the PGPS9K datasets. The Geometry3K is also provided in this database.

Note: Due to copyright restrictions, we are currently only providing links for these datasets. Full datasets for tuning and evaluation organized by us will be provided via email. If you need it, please contact us by email.

For more details, please refer to our paper and GitHub repository. If you find our work helpful, please consider starring ⭐ in this repository and citing us:

@article{xia2024geox,
  title={GeoX: Geometric Problem Solving Through Unified Formalized Vision-Language Pre-training},
  author={Xia, Renqiu and Li, Mingsheng and Ye, Hancheng and Wu, Wenjie and Zhou, Hongbin and Yuan, Jiakang and Peng, Tianshuo and Cai, Xinyu and Yan, Xiangchao and Wang, Bin and others},
  journal={arXiv preprint arXiv:2412.11863},
  year={2024}
}