SGOCR: Spatially-Grounded OCR Visual Question Answering Dataset

SGOCR is a dataset for training and evaluating models that must read, localize, and reason about text in images. It contains grounded question-answer pairs over ChartQA, TextOCR, and COCO/COCO-Text source images. The dataset was created by author dreeseaw and was last updated on 2026-04-18.

Use Cases

• Training OCR-aware VQA models based on grounded question-answer pairs.
• Evaluating text grounding and localization performance based on spatially-referenced annotations.
• Conducting data-centric experiments for scene text understanding based on multiple source image collections.
• Developing models for region-conditioned question answering based on the dataset's design.

Strengths

• Designed for a specific, complex task combining OCR, localization, and reasoning.
• Integrates data from multiple established source datasets: ChartQA, TextOCR, and COCO/COCO-Text.
• Last updated on 2026-04-18, indicating recent maintenance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and dataset size are unknown, which may limit suitability assessment.
• Data may reflect biases inherent to the source image collections used.

Provenance

Source

Hugging Face dataset created by dreeseaw.

Collection Method

Contains grounded question-answer pairs over ChartQA, TextOCR, and COCO/COCO-Text source images.

Time Range

null

Freshness

Last updated 2026-04-18 22:19:22.

Geography

null