| # GeoPeP: Geometric Perception for Mathematical Reasoning (diagram-caption subset) | |
| Welcome to the caption-based subset of GeoPeP (Geomatic PercePtion-oriented dataset) — a large-scale dataset designed to enhance geometric perception and visual reasoning capabilities in multimodal large language models (MLLMs). | |
| 📘 **Overview** | |
| GeoPeP-Caption consists of 100,000 synthetic diagrams paired with automatically generated fine-grained captions. Each caption provides a structured description of the geometric primitives within the diagram, enabling vision-language models to learn detailed spatial reasoning and accurate object grounding. | |
| This dataset is part of the broader GeoPeP project, which includes both caption-based (will be released soon) and instruction-following subsets to bridge the gap between visual perception and mathematical reasoning. | |
| 🔗 **Download Link** | |
| You can access and download the GeoPeP-Caption dataset here: | |
| 👉 [GeoPeP-Caption](https://drive.google.com/file/d/1YPXaNhY2O5fSUsZYQjJ5vV5oKkSTrymP/view?usp=drive_link) | |
| 📦 **Dataset Structure** | |
| ``` | |
| { | |
| "image": "image_id.png", | |
| "conversations": [ | |
| { | |
| "from": "human", | |
| "value": "<image>\nPlease provide structured geometric information for the image." | |
| }, | |
| { | |
| "from": "gpt", | |
| "value": "First, I count 3 prominent objects in the image, including 1 equilateral quadrilateral, 1 isosceles trapezoid, 1 pentagon. Next, for the shape information, object BFLV is an equilateral quadrilateral, object CGOY is an isosceles trapezoid, object AHIWX is a pentagon. Furthermore, I also know the fine-grained bounding box coordinates: the equilateral quadrilateral BFLV is located at [0.32, 0.46, 0.56, 0.68], the isosceles trapezoid CGOY is located at [0.23, 0.29, 0.58, 0.61], the pentagon AHIWX is located at [0.22, 0.24, 0.78, 0.76]. Finally, let me explain the relationships: equilateral quadrilateral BFLV is in the bottom right relative to isosceles trapezoid CGOY; equilateral quadrilateral BFLV is in the bottom left relative to pentagon AHIWX." | |
| } | |
| ] | |
| } | |
| ``` | |
| 🤖 **Intended Use** | |
| * Pretraining and fine-tuning Multimodal LLMs. | |
| * Enhancing geometric perception and caption grounding in visual models. | |
| * Supporting research in visual reasoning and math understanding | |
| 📚 **Citation** | |
| If you use this dataset, please cite our work: | |
| ``` | |
| @article{sun2025mathglance, | |
| author = {Sun, Yanpeng and Zhang, Shan and Tang, Wei and Chen, Aotian and Koniusz, Piotr and Zou, Kai and Xue, Yuan and van den Hengel, Anton}, | |
| title = {MATHGLANCE: Multimodal Large Language Models Do Not Know Where to Look in Mathematical Diagrams}, | |
| journal = {arXiv preprint arXiv:2503.20745}, | |
| year = {2025} | |
| } | |
| ``` | |
| 📬 **Contact** | |
| If you have any questions or suggestions, please open an issue or contact [email protected]. | |