pubmed-ocr

# PubMed-OCR: PMC Open Access OCR Annotations PubMed-OCR is an OCR-centric corpus of scientific articles derived from PubMed Central Open Access PDFs. Each **page** is rendered to an image and annotated with **Google Cloud Vision OCR**, released in a compact JSON schema with **word-, line-, and paragraph-level** bounding boxes. **Scale (release):** - **209.5K** articles - **~1.5M** pages - **~1.3B** words (OCR tokens) This dataset is intended to support layout-aware modeling, coordinate-grounded QA, and evaluation of OCR-dependent pipelines on scientific documents. ## Dataset Details ### Dataset Description - **Curated by:** Roots.ai - **Point of contact:** [e-mail] - **Language:** English (primarily; see limitations) - **Data unit:** **1 row = 1 PDF page** (unique by `{basename, page}`) - **License:** See **Licensing** section (source-article licenses; per-row `license` field) ### Dataset Sources - **Repository:** https://huggingface.co/datasets/rootsautomation/pubmed-ocr - **Paper:** [PubMed-OCR: PMC Open Access OCR Annotations](https://huggingface.co/papers/2601.11425) - **Source corpus:** PubMed Central Open Access (PMCOA) ## Uses ### Direct Use PubMed-OCR is suitable for: - Training/evaluating **OCR-aware** or **layout-aware** document models - Testing robustness of pipelines that depend on OCR (parsing, retrieval, extraction) - Building tasks that require **coordinate-grounded evidence** (e.g., quote-and-locate, region attribution) - Benchmark curation for scientific PDFs (tables, formulas, captions, references) ### Out-of-Scope Use - Do **not** treat OCR output as gold text; it contains recognition errors. - Not intended for clinical/medical decision-making. - Not intended for learning copyrighted content outside the applicable license terms. - Not intended as a reading-order ground truth dataset. ## Dataset Structure ### Data Instances Each row corresponds to a single page. Key identifiers: - `basename`: page group identifier (article-level) - `page`: page index within the article `ocr_json` is a JSON string containing OCR outputs with bounding boxes in **pixel coordinates** for the rendered page image. Example (schema sketch; fields may include additional metadata): ```json { "image": {"width": 1275, "height": 1650, "dpi": 150}, "text": { "words": [{"text": "Introduction", "bbox": [74, 132, 210, 156]}], "lines": [{"text": "Introduction", "bbox": [74, 130, 612, 160]}], "paragraphs": [{"text": "…", "bbox": [70, 120, 1180, 420]}] } } ``` ### Data Fields * `basename` *(string)*: article/page group identifier. * `page` *(int32)*: page index within the PDF/article. * `license` *(string)*: the **source article’s license** (e.g., `cc-by-4.0`, `cc-by-nc-4.0`, …). * `pmid` *(string)*: PubMed ID when available. * `accession_id` *(string)*: accession identifier (e.g., PMCID or internal ID). * `article_citation` *(string)*: a citation string for the source article. * `pdf_bytes` *(binary)*: raw PDF bytes **when redistribution is permitted**; may be empty/null otherwise. * `ocr_json` *(string)*: OCR output JSON (see above). ### Splits This release is provided as a single split (`train`) because it is primarily a **corpus**. For benchmarking, consider constructing evaluation splits that reduce leakage, e.g.: * **Journal-level splits** (hold out entire journals) * **Time-based splits** (hold out by publication year) * **PMID/PMCID disjoint splits** (article-level separation) ## Dataset Creation ### Curation Rationale Scientific PDFs are dense (formulas, tables, multi-column layouts). Many PMCOA datasets rely on PDF/XML alignment, which can miss scanned pages or inherit parser noise. PubMed-OCR provides OCR-native supervision directly from rendered page images, enabling OCR-dependent evaluation and layout-aware learning without PDF/XML alignment. ### Source Data #### Data Collection and Processing High-level pipeline: 1. Download PubMed Central Open Access PDFs (PMCOA) and filter to licenses permitting redistribution of derived artifacts. 2. Uniformly sample 209.5K documents. 3. Render each page at **150 DPI**. 4. Run **Google Cloud Vision** `document_text_detection` on page images. 5. Extract word- and paragraph-level polygons and canonicalize to axis-aligned bboxes `[x1, y1, x3, y3]`. 6. Reconstruct **line** bboxes by clustering words with similar vertical alignment (heuristic). 7. Emit one row per page with `ocr_json` (+ `pdf_bytes` where permitted). #### Who are the source data producers? The source texts were authored by scientific article authors and published via journals hosted in PubMed Central Open Access. ### Annotations #### Annotation process Annotations are machine-generated via Google Cloud Vision OCR. * **Words / paragraphs:** provided by the OCR engine * **Lines:** reconstructed heuristically from word boxes (see Limitations) #### Who are the annotators? The OCR engine is the annotator. No manual annotation was performed in this release. #### Personal and Sensitive Information Scientific articles can contain author names, affiliations, acknowledgements, emails, and citations. Content is drawn from publicly available PMCOA articles; no additional anonymization is applied. ## Bias, Risks, and Limitations * **Single OCR engine:** outputs reflect Google Vision’s strengths/weaknesses and may not generalize to other OCR systems. * **Heuristic line reconstruction:** line grouping and reading order can be imperfect, especially in multi-column layouts and around formulas/tables. * **Axis-aligned boxes:** original OCR polygons are simplified to rectangles. * **Domain skew:** PMCOA’s journal distribution is heavy-tailed (high-volume journals dominate). * **Non-text regions:** this dataset does not provide gold structure for tables/figures/formulas (only what OCR emits + derived lines). ### Recommendations * When reporting results, specify whether you use **words**, **lines**, or **paragraphs**, and whether you re-linearize text. * For fair evaluation, prefer **journal-disjoint** or **article-disjoint** splits. * If you need table/figure structure, pair this with a layout/table dataset (or run a layout model on top). ## Licensing This dataset contains content derived from PMCOA articles. * Each example inherits the **license of its source article**, recorded in the `license` field. * Users are responsible for complying with the license terms for any subset they use. * If `pdf_bytes` is present, it is provided only where redistribution is permitted. ## Citation If you use PubMed-OCR, please cite: ```bibtex @article{heidenreich2025pubmedocr, title={PubMed-OCR: PMC Open Access OCR Annotations}, author={Heidenreich, Hunter and Getachew, Yosheb and Dinica, Olivia and Elliott, Ben}, journal={arXiv preprint arXiv:2601.11425}, year={2025} } ``` ## How to Load ```python from datasets import load_dataset import json ds = load_dataset("rootsautomation/pubmed-ocr", split="train") row = ds[0] ocr = json.loads(row["ocr_json"]) words = ocr["text"]["words"] ``` For large-scale iteration, consider streaming: ```python ds = load_dataset("rootsautomation/pubmed-ocr", split="train", streaming=True) for row in ds: ocr = json.loads(row["ocr_json"]) ... ``` ## Models Trained on this Data - [GutenOCR-3B](https://huggingface.co/rootsautomation/GutenOCR-3B) - [GutenOCR-7B](https://huggingface.co/rootsautomation/GutenOCR-7B)