209,500 open-access scientific articles from PubMed Central were rendered to images and processed with Google Cloud Vision OCR. The corpus contains approximately 1.5 million pages and 1.3 billion OCR tokens. It is released by rootsautomation in a JSON schema with bounding boxes at word, line, and paragraph levels.
Use Cases
- Layout-aware modeling based on bounding box annotations for words, lines, and paragraphs.
- Coordinate-grounded question answering based on OCR-derived text and spatial information.
- Evaluation of OCR and document understanding systems based on a large-scale corpus of scientific pages.
Strengths
- Large scale with 209,500 articles and approximately 1.5 million pages.
- High token count of approximately 1.3 billion OCR words.
- Multi-level annotation schema providing bounding boxes for words, lines, and paragraphs.
Limitations
- Description metadata is limited; actual data quality requires manual inspection after download.
- Column-level documentation is absent; field semantics must be inferred after download.
- Last updated 2026-01-22 19:58:29; freshness should be verified.
Provenance
- Source
- PubMed Central Open Access PDFs.
- Collection Method
- Pages rendered to images and annotated with Google Cloud Vision OCR.
- Freshness
- 2026-01-22 19:58:29