GeoGPT4V: Towards Geometric Multi-modal Large Language Models with Geometric Image Generation

This repository contains the code and data for the paper titled “GeoGPT4V: Towards Geometric Multi-modal Large Language Models with Geometric Image Generation”.

Contents

• Install
• Usage
• Train
• Dataset
• Model

Install

1. Clone this repository.

2. Install Package.

   conda create -n geogpt4v python=3.10 -y
   conda activate geogpt4v
   pip install -r requirements.txt

3. Install Wolfarm engine. Please follow its offical tutorial.

Usage

Data Preparation

1. Download following open-source datasets or use your own datasets:

   1. Geometry3K
   2. GeoQA
   3. UniGeo

2. Transform the dataset into the following format and save it as a jsonline file:

   # multi-choice quetion example
   {"id": 1, "question": "For the pair of similar figures, find the area of the green figure.", "choices": ["20.4", "28.6", "56.0", "78.4"], "answer": "D", "image": "image path"}
   
   # open-ending quetsion example
   {"id": 2, "question": "Prove that △ABC is congruent to △DEF.", "answer": "Because AB = DE, BC = EF, and ∠ABC is equal to ∠DEF, △ABC is congruent to △DEF.", "image": "image path"}

3. [optional] If you are using your own dataset, please modify the function ‘construct_prompt’ in ‘./pipeline/gen_instruction_mp.py’.

Data Generation

Run the scripts in the following order, remember to modify the dataset path and API key in the scripts.

1. sh scripts/gen_instruction_gpt4v_mp.sh
2. sh scripts/gen_image_mp.sh
3. sh scripts/rerank_gp4v_mp.sh
4. sh scripts/filter.sh

Train

You can use following models’ offical training code or use your own code. You can run scripts/convert_format.sh to transform the GeoGPT4V dataset into the required format for model training.

1. LLaVA
2. ShareGPT4V
3. InternVL-Chat

Dataset

You can download the datasets from the following links.

Dataset	Link	Note
GeoGPT4V-1.0	https://huggingface.co/datasets/caishihao/GeoGPT4V-1.0	The dataset we used in the paper.
GeoGPT4V-1.1	https://huggingface.co/datasets/caishihao/GeoGPT4V-1.1	The dataset after applying the rule-based (image size) filtering.

Model

You can download the models from the following links.

Model	Link
LLaVA-1.5-7B-GeoGPT4V	https://huggingface.co/caishihao/GeoGPT4V-LLaVA-1.5-7B-v1
LLaVA-1.5-13B-GeoGPT4V	https://huggingface.co/caishihao/GeoGPT4V-LLaVA-1.5-13B-v1
ShareGPT4V-7B-GeoGPT4V	https://huggingface.co/caishihao/GeoGPT4V-ShareGPT4V-7B-v1
ShareGPT4V-1.5-13B-GeoGPT4V	https://huggingface.co/caishihao/GeoGPT4V-ShareGPT4V-13B-v1
InternVL-Chat-V1.2-Plus-40B-GeoGPT4V	https://huggingface.co/Rosiness/GeoGPT4V-InternVL-Chat-40B