Mathematical Expression Recognition Dataset with Over One Million Training Instances

UniMER is a dataset for advancing Mathematical Expression Recognition (MER), containing over one million training instances of diverse mathematical expressions. It was created by author 'wanderkid' and includes a separate test set for evaluation. The dataset was last updated on the Hugging Face platform in March 2025.

Use Cases

• Train image-to-text models to recognize and transcribe complex mathematical symbols and structures from the provided expression images.
• Benchmark the performance of MER systems on the UniMER Test Set to evaluate generalization across diverse expression types.
• Fine-tune pre-trained vision-language models using the labeled image-text pairs for specialized mathematical content understanding.
• Develop and evaluate models for parsing the spatial layout and relationships between components in mathematical expression images.

Strengths

• Contains over 1,000,000 training instances, providing a large-scale resource for model training.
• Includes a separate, meticulously curated test set for reliable model evaluation.

Limitations

• Specific details on image resolution, annotation quality, and class distribution are not provided in the input.
• The geographic and temporal origins of the source expressions are unknown, which may introduce bias.

Provenance

Source

Hugging Face dataset uploaded by user 'wanderkid'.

Collection Method

Curated collection of mathematical expression images and their corresponding text representations.

Freshness

Dataset was last updated on 2025-03-25.