Incinciblecolonel/MathStrike

Math-Strike Dataset

Dataset Description

Math-Strike is a novel dataset featuring real and synthetic strike-outs, component-level annotations, and aligned LaTeX, designed to support research on strike-out removal and handwritten mathematical formula recognition. Handwritten STEM manuscripts often contain strike-outs, overwrites, and other markings that disrupt mathematical content structure and reduce OCR/VLM recognition accuracy. This dataset provides a rigorous benchmark for training and evaluating models on connected-component classification, geometry-aware inpainting, and downstream compilable LaTeX generation.

The collection originates from the 2024-2025 mid-term and final examinations of a university's Multivariate Calculus Elite class.

Dataset Subsets

The dataset is divided into two main subsets, easily accessible via Hugging Face's configuration system:

• real (Math-Strike-Real): Contains only real, authentic strike-outs collected directly from student exam papers.
• synth (Math-Strike-Synth): Contains synthetic strike-outs generated using an augmentation taxonomy of 20 diverse strike-out styles (e.g., wavy, zigzag, multi-crossed, blackout) to address class imbalance and improve model robustness against complex occlusion patterns. Additionally, the original_img folder is provided in the repository, containing the full page-level raw scans with minimal border cropping and all original marks.

Dataset Structure

Data Instances

When loading the dataset via the datasets library, each instance represents a single connected component extracted from the manuscript.

Data Fields

Field	Type	Description
id	int	Global component identifier.
page_img	string	Source page filename.
bbox	list	Bounding box coordinates in pixels.
centroid	list	Geometric centre (x, y) of the component.
box_area / pixel_area	int	Foreground-pixel count within the component.
label	int	0 = retained (valid mathematical symbol), 1 = removed (strike-out/noise).
comp_size	list	Height and width of the bounding box.
image	Image	The actual cropped PNG image of the component.
file	string	The actual path of the image.

Usage

You can easily load the dataset using the Hugging Face datasets library. The images and metadata are already pre-aligned in Parquet format.

from datasets import load_dataset

# Load the subset with real strike-outs
dataset_real = load_dataset("Incinciblecolonel/MathStrike", name="real")

# Load the subset with synthetic strike-outs
dataset_synth = load_dataset("Incinciblecolonel/MathStrike", name="synth")

# Access the training split
train_data = dataset_real["train"]
print(train_data[0]) 

# Directly visualize the component image
train_data[0]["image"].show()

Alternative Download (Raw Files)

For researchers who prefer downloading the raw image folders and JSON annotations directly instead of using the datasets library, the complete archived dataset (ZIP) is hosted on Kaggle:

👉 Download Math-Strike on Kaggle

License

The Math-Strike dataset is distributed under the CC-BY-NC 4.0 license for non-commercial research and educational use.

Citation

If you find this dataset useful for your research, please cite our paper:

@article{shen2026mathscrub,
  title={MathScrub: Single-Pipeline Strike-Out Removal and LaTeX Generation for Handwritten Calculus Expressions},
  author={Shen, Sen and Liu, Ning and Song, Lintao and Han, Dongkun},
  journal={IEEE Transactions on Learning Technologies},
  year={2026},
  publisher={IEEE}
}