TriGeoBench

TriGeoBench is a geometry-centric multimodal mathematics benchmark designed for evaluating mathematical reasoning with visual diagrams and image-to-TikZ generation. The dataset contains mathematical problems, solutions, diagram images, and corresponding TikZ annotations.

This repository is anonymized for peer review. Author and institution information will be added upon acceptance.

Dataset Files

The dataset contains four Parquet files:

TriGeoBench
├── README.md
├── image2tikz/
│   ├── train.parquet
│   └── test.parquet
└── question/
    ├── train.parquet
    └── test.parquet

The dataset supports two tasks:

1. Image-to-TikZ generation: generating TikZ code from a geometric diagram image.
2. Multimodal mathematical reasoning: solving math problems with textual questions, solutions, and associated figures.

Image-to-TikZ Data

Files:

image2tikz/train.parquet
image2tikz/test.parquet

Each row corresponds to one diagram image and its ground-truth TikZ code.

Fields

Field	Description
key	Unique figure identifier. It is composed of <problem_id>_<position>_<figure_index>, where position indicates whether the figure appears in the question or the solution. This key can be linked to the corresponding problem in the question-level data.
image	Base64-encoded image.
latex_gt	Ground-truth TikZ code corresponding to the image.
difficulty	Figure complexity level. Possible values are 容易, 中等, and 困难.

Question-Level Data

Files:

question/train.parquet
question/test.parquet

Each row corresponds to one mathematical problem, including the problem text, solution, metadata, and associated figures.

Fields

Field	Description
sample_id	Unique problem identifier. It can be linked to the key field in the image-to-TikZ data.
difficulty	Problem difficulty level. Possible values are 容易, 中等, and 困难.
question_type	Problem type. Possible values include 选择题, 填空题, 解答题, and 证明题.
knowledge_point	Main mathematical knowledge area. Possible values include 向量, 函数, 平面几何, 立体几何, and 解析几何.
question	Problem statement in LaTeX format.
solution	Solution or answer in LaTeX format.
q_figX	Base64-encoded image of the X-th figure appearing in the question.
q_figX_latex_gt	Ground-truth TikZ code of the X-th question figure.
s_figY	Base64-encoded image of the Y-th figure appearing in the solution.
s_figY_latex_gt	Ground-truth TikZ code of the Y-th solution figure.

Here, X and Y denote figure indices. A problem may contain different numbers of question-side and solution-side figures.

Data Splits

The dataset is split into training and test sets for both tasks:

Task	Train File	Test File
Image-to-TikZ	image2tikz_train.parquet	image2tikz_test.parquet
Mathematical Reasoning	question_train.parquet	question_test.parquet

Loading the Dataset

The Parquet files can be loaded with pandas:

import pandas as pd

image2tikz_train = pd.read_parquet("image2tikz_train.parquet")
image2tikz_test = pd.read_parquet("image2tikz_test.parquet")

question_train = pd.read_parquet("question_train.parquet")
question_test = pd.read_parquet("question_test.parquet")

Base64-encoded images can be decoded as follows:

import base64
from io import BytesIO
from PIL import Image

def decode_base64_image(image_base64):
    image_bytes = base64.b64decode(image_base64)
    return Image.open(BytesIO(image_bytes)).convert("RGB")

img = decode_base64_image(image2tikz_train.iloc[0]["image"])
img.show()

Intended Use

TriGeoBench is intended for research on:

• multimodal mathematical reasoning;
• geometry-centric visual question answering;
• image-to-TikZ generation;
• evaluating whether models can reason over precise geometric structures;
• studying the interaction between textual math problems, visual diagrams, and symbolic diagram representations.

Limitations

The dataset focuses on geometry-centric middle- and high-school mathematics problems. The annotations include LaTeX-formatted problem texts and TikZ code for figures. Although the dataset has been processed and checked, residual annotation errors may remain.

Anonymous Review Notice

This repository is anonymized for peer review. Please do not attempt to identify the authors during the review process.