PubTabNet Table Images with HTML Annotations

## Overview PubTabNet is a large-scale dataset for image-based table recognition containing 568,000+ images of tabular data, each annotated with the corresponding HTML representation describing both table structure and cell content. The tables were extracted from scientific publications in the PubMed Central Open Access Subset (commercial use collection). The dataset is distributed in WebDataset format and combines image and text modalities, making it suitable for training table structure recognition, OCR, and document understanding models. ## Schema - `image` — image (PNG) — rendered table region cropped from the source PDF - `html` / `structure` — text — HTML annotation encoding table structure (rows, columns, spans) and cell text - `filename` — string — original identifier tying back to the source PMC article - `split` — string — train/val/test partition (per original PubTabNet release) ## Sources - HuggingFace: https://huggingface.co/datasets/ajimeno/PubTabNet — license: CDLA-Sharing-1.0 - Original dataset: IBM Research PubTabNet (https://github.com/ibm-aur-nlp/PubTabNet) - Underlying corpus: PubMed Central Open Access Subset (commercial use collection) - Reference: Zhong et al., "Image-based table recognition: data, model, and evaluation" (arXiv:1911.10683) ## Methodology Table regions were identified by aligning the PDF rendering of each article with its corresponding XML representation in the PubMed Central Open Access Subset. Where the PDF layout could be reliably matched to XML-defined table content, the table region was cropped as an image and paired with HTML reconstructed from the XML structure, including cell-level text and span information. This HuggingFace mirror repackages the original IBM release into the WebDataset format for streamable consumption. ## Known gaps & limitations Content is restricted to scientific/biomedical publications, so visual style, terminology, and table conventions are skewed toward that domain — generalization to financial, legal, or handwritten tables is not guaranteed. Annotations are derived from PDF/XML alignment heuristics and may contain misaligned cells or imperfect HTML for complex tables (deeply nested headers, rotated text, multi-page tables). Cell content text is the XML-extracted string, not OCR — it may differ from what a model would read off pixels. Source does not enumerate per-example quality metadata; buyers should validate empirically on their target distribution. ## Intended use & out-of-scope - IS for: training and evaluating table structure recognition models, document AI / layout models, image-to-HTML sequence models, OCR pipelines that need table-aware post-processing, and RAG over scientific documents. - NOT for: claims of coverage outside biomedical literature; benchmark training without checking leakage against PubTabNet's official val/test splits; tasks needing handwritten or non-English tables. ## License CDLA-Sharing-1.0 (Community Data License Agreement – Sharing). Share-alike: redistribution and derivative datasets are permitted provided the same license terms are preserved. _Federated dataset: 10 parquet shards, 4.64 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: PubTabNet — 568K Table Images with HTML Annotations 568K+ scientific table images paired with HTML structure annotations, extracted from PubMed Central Open Access articles. Standard benchmark for image-based table recognition and document AI.