MM-Math-Align

Hard Negative Contrastive Learning for Fine-Grained Geometric Understanding in Large Multimodal Models

| [🐙 Github Code](https://github.com/THU-KEG/MMGeoLM) |
[📃 Paper](https://arxiv.org/abs/2505.20152) |

# Dataset description:

We release **MM-Math-Align**, a dataset built upon [**MM-Math**](https://huggingface.co/datasets/THU-KEG/MM_Math), which is derived from actual geometry questions used in middle school exams. Each sample contains the original geometric diagram(**original_image**), a Python script's image(**positive_image**) that approximately reconstructs the diagram, a caption(**positive_caption**) describing the positive image, 10 negative Python script images(**negative_idx_image**) generated based on the positive image, and 10 corresponding negative captions(**negative_idx_caption**). The dataset consists of a total of 4,021 samples.

# Citation
```
@misc{sun2025hardnegativecontrastivelearning,
title={Hard Negative Contrastive Learning for Fine-Grained Geometric Understanding in Large Multimodal Models},
author={Kai Sun and Yushi Bai and Zhen Yang and Jiajie Zhang and Ji Qi and Lei Hou and Juanzi Li},
year={2025},
eprint={2505.20152},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2505.20152},
}
```

# 面向大型多模态模型细粒度几何理解的难样本对比学习

| [🐙 Github 代码](https://github.com/THU-KEG/MMGeoLM) |
[📃 论文](https://arxiv.org/abs/2505.20152) |

# 数据集说明：

我们发布了**MM-Math-Align**数据集，该数据集基于[**MM-Math**](https://huggingface.co/datasets/THU-KEG/MM_Math)构建，其原始数据来源于中学考试中的真实几何考题。每个样本包含原始几何图形图（original_image）、一张可近似重建该图形的Python脚本生成图像（positive_image）、一段描述该正样本图像的说明文本（positive_caption）、10张基于正样本图像生成的负样本Python脚本图像（negative_idx_image），以及10条对应的负样本说明文本（negative_idx_caption）。该数据集总计包含4021个样本。

# 引用

@misc{sun2025hardnegativecontrastivelearning,
title={面向大型多模态模型细粒度几何理解的难样本对比学习},
author={Kai Sun and Yushi Bai and Zhen Yang and Jiajie Zhang and Ji Qi and Lei Hou and Juanzi Li},
year={2025},
eprint={2505.20152},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2505.20152},
}