MNIST-VQA: Synthetic Visual Question Answering Dataset with MNIST Digits

MNIST-VQA is a synthetic Visual Question Answering dataset generated from MNIST digits placed on a 3x3 grid. It is designed to test spatial reasoning, object localization, counting, and existence verification capabilities of VQA models. The dataset was created by author star092304 and was last updated on the Hugging Face platform in June 2026.

Use Cases

• Benchmarking VQA model performance on spatial reasoning tasks based on the 3x3 grid layout.
• Training models for object localization and counting based on the presence of MNIST digits.
• Evaluating model capabilities for existence verification questions based on the dataset's design.
• Testing model generalization across difficulty levels based on the density constraint variations (2-4 digits).

Strengths

• Dataset is explicitly designed for testing specific VQA capabilities: spatial reasoning, object localization, counting, and existence verification.
• Offers structured difficulty levels, with variations characterized by the number of digits present per image (Density Constraint).
• Built upon the well-known MNIST digit dataset, providing a familiar and standardized visual component.

Limitations

• Description metadata is limited; actual data quality, column definitions, and file formats require manual inspection after download.
• Row count, file size, and license information are unknown, which may limit suitability assessment.
• Data is synthetic, which may not fully capture the complexity and noise of real-world visual question answering scenarios.

Provenance

Source

huggingface

Collection Method

Synthetically generated from the MNIST dataset.

Freshness

Last updated 2026-06-19 10:07:51; freshness should be verified.