lschiemanowski/scannedtables

Scanned tables

Crops of printed tables from 215 public-sector, scientific and historical documents, digital and scanned, with a transcription of the table as HTML under a fixed format and the table as structured JSON. Validation and test are fully transcribed; most of train is not yet. Schema version 0.1.

Split	Crop parts	Accepted transcriptions	Rejected	Unlabeled
train	8,786	958	9	7,819
validation	934	933	1	0
test	975	974	1	0

Where the crops come from

The crops are the table regions of the frozen v1 splits of a local table OCR benchmark built from 215 source PDFs. The splits are disjoint by publication series. Each region was recropped by the table reader's document preparation: pages rendered at 300 DPI with pdfium, the detected table box extended upward over the table's title and units line and downward over its footnotes, with layout detection (Docling Heron) stopping the extension at running text. Tables printed sideways are rotated upright. A table split across pages has one row per part.

406 crop parts are held out. A crop-level rights and privacy review held named compensation, pension and reimbursement figures, personal declarations, signatures, contractor-authored or third-party material, and components with conflicting or unresolved reuse terms. A second screen of the transcriptions held every table that lists named individuals (boards, committees, ministers, fees, donors, related parties) and checked each mention of copyright, permissions or personal identifiers in its context. Both screens work mainly from text and can miss material that neither caught; please report anything that should not be here.

Labels

• transcription reproduces what is printed, in the format described in transcription-format.md: a # Title line when a title is printed, an HTML table with thead for printed column headings, data-level for the printed indentation of row labels, <sup> for footnote markers and tfoot for footnote lines. A crop without a table is NOT_A_TABLE.
• tables is the JSON (format 0.3, table-format.md, table.schema.json) that the table reader's deterministic converter proposes from the transcription. Printed values stay strings.
• transcription_model names the model that wrote the transcription. Current labels were written by GPT-6.1 Terra and have not been reviewed by a person. transcription_status is accepted when the transcription parses into tables without blocking issues and was not cut off, rejected otherwise, and unlabeled where no transcription exists yet.
• v1_label is the original benchmark's classification of the crop: table, wrongly_cropped or not_a_table. It refers to the original tighter crop.

Columns

region_id (crop key and part), table_id, part, split, image (PNG, upright), image_sha256, rotation_cw (clockwise rotation applied to the page region), document_id, source_id, medium (scanned or born_digital), page (one-based PDF page), crop_box and page_size (PDF points, before rotation), v1_label, transcription, transcription_status, transcription_model, region_status (tables, not_a_table or failed), tables (JSON text), rights_basis, attribution, source_url.

Built and used by tablereader; the model trained on it is lschiemanowski/tablereader-lightonocr-v0.

Rights

Source terms vary by document; see LICENSE and documents.jsonl, which gives each document's rights basis, source link and attribution. Cropping and transcription are adaptations of the source documents.

The annotations (transcription, tables and the label columns) are released under the Creative Commons Attribution 4.0 International license (CC BY 4.0, https://creativecommons.org/licenses/by/4.0/). The crop images remain under their source documents' terms.