ai-historian/german-newspaper-tables

📋 On the licensing of this data

Every item here is labelled to the best of my ability. The rights statement is taken per issue from what the holding institution recorded in the Deutsche Digitale Bibliothek, captured at download time — never inferred, and never applied at newspaper level to issues that may differ.

Labelling at this scale is imperfect. If you believe anything here is mislabelled, or you hold rights in any of this material, please write to lorenz.hufe@posteo.de — I will resolve it immediately.

German Newspaper Tables (1710–1945)

1,044,813 table images cut from German-language newspapers, each paired with the OCR transcription of that table and full provenance metadata. All items are in the public domain.

Historical tables are the hard case for document layout and OCR systems: the formats vary enormously across two and a half centuries, from narrow columns of exchange rates to full-page statistical grids. This dataset exists to make that variety available at scale.

Samples	1,044,813
Source pages	9,160,180 newspaper pages screened
Issues	320,626
Newspapers	1,350
Places of publication	157
Years	1710–1945
Size	392.6 GB, 1,024 WebDataset shards
Licence	CC Public Domain Mark 1.0 (1,044,076) · CC0 1.0 (737)

Format

WebDataset. Each sample is three files sharing one key:

<decade>/<issue_id>_<page>_<region_order>.jpg    the table crop
<decade>/<issue_id>_<page>_<region_order>.txt    OCR transcription, one line per table line
<decade>/<issue_id>_<page>_<region_order>.json   metadata

import webdataset as wds

ds = wds.WebDataset("https://huggingface.co/datasets/ai-historian/german-newspaper-tables/resolve/main/tables-{2A,2B}.tar")
for sample in ds:
    image = sample["jpg"]
    text  = sample["txt"].decode()
    meta  = json.loads(sample["json"])

Or with datasets:

from datasets import load_dataset
ds = load_dataset("ai-historian/german-newspaper-tables", split="train", streaming=True)

Metadata fields: key, year, date, decade, issue_id, page, region_order, newspaper, place, zdb_id, license, bbox_on_page, page_size, crop_size, layout_score, n_table_lines, n_tables_on_page, source_image_url, ddb_url.

bbox_on_page locates the table on the full scan, and ddb_url links to the digitised page in the Deutsche Digitale Bibliothek — so every crop can be traced back to its source.

How it was made

Source images come from the Deutsches Zeitungsportal of the Deutsche Digitale Bibliothek. Every page was processed with a two-stage OCR pipeline: layout detection with PP-DocLayout_plus-L, then transcription of each detected region with churro-3B. The table regions of that run are what this dataset contains — no separate table detector was run.

From 8,178,018 detected table regions across all 9.16 M pages, samples were kept when they passed all of: layout confidence ≥ 0.70, at least 6 transcribed lines, both sides ≥ 400 px, and 2–80 % of the page area. That leaves 1,512,255; of those, the 1,044,813 whose issue carries a public-domain rights statement are published here. Crops are the region box plus 2 % padding, capped at 2000 px wide, JPEG q88. Image bytes are copied unchanged from the production run.

Known limitations

Precision is about 97 %. Measured by reviewing 172 crops by eye across two independent samples. So expect on the order of 31,000 items that are not tables — almost always classified-advertisement pages, whose column rules read to the model like table rules. Filter harder on layout_score and n_table_lines if you need cleaner data.

A region is a detection, not a printed table. One printed table can fall into several boxes, and a page genuinely carrying several tables produces several samples. Table structure (rows, columns, cells) is not annotated — this is detection output plus reading-order text, not a table-structure ground truth.

OCR text is machine output, not corrected. churro-3B on Fraktur and historical orthography. Known failure modes include occasional repetition loops and truncation on very dense regions, and rare hallucinated modern boilerplate. Orthography is unstable across the period (ſ/s, ß/ss, ae/ä); normalise before matching.

Coverage follows what was digitised, not what was printed. The corpus is licence-shaped and regionally uneven: strong in the west and Saxony, with Berlin effectively absent before 1871. Counts per decade below are a property of the holdings, not of the press.

Distribution by decade

decade	tables	share
1710s	10	0.0 %
1720s	22	0.0 %
1730s	70	0.0 %
1740s	8	0.0 %
1750s	457	0.0 %
1760s	298	0.0 %
1770s	1,530	0.1 %
1780s	2,591	0.2 %
1790s	1,844	0.2 %
1800s	2,571	0.2 %
1810s	3,373	0.3 %
1820s	4,706	0.5 %
1830s	15,019	1.4 %
1840s	25,613	2.5 %
1850s	59,821	5.7 %
1860s	60,828	5.8 %
1870s	123,217	11.8 %
1880s	165,516	15.8 %
1890s	175,892	16.8 %
1900s	163,069	15.6 %
1910s	81,332	7.8 %
1920s	94,561	9.1 %
1930s	48,486	4.6 %
1940s	13,979	1.3 %
total	1,044,813	100 %

Tables also became far more common over time: from roughly 1 table per 1,600 pages in the 1710s to 1 per 4 pages in the 1900s, a ~400-fold rise in density.

Content warning

This is an unfiltered slice of the German press between 1710 and 1945, and it includes material from the National Socialist period. Among the administrative tables of the Deutscher Reichsanzeiger are expatriation lists naming Jewish citizens stripped of their citizenship, as well as ration-card allocations and similar instruments of the Nazi state.

These are genuine historical records and are published here for research. They are not neutral illustration material, and they name real, persecuted individuals. Please consider this before reusing individual items, and particularly before presenting them out of context.

Licence and attribution

Digitisation rights statements are per issue, taken from the issue's own manifest at download time: CC Public Domain Mark 1.0 or CC0 1.0. Items with "In Copyright", "Copyright Not Evaluated" or unresolved rights were excluded from this release and are not published.

When using this dataset, please credit the Deutsches Zeitungsportal / Deutsche Digitale Bibliothek and the holding libraries as the source of the digitised newspapers.

Citation

@misc{german_newspaper_tables_2026,
  title  = {German Newspaper Tables (1710--1945)},
  author = {Hufe, Lorenz},
  year   = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/ai-historian/german-newspaper-tables}},
  note   = {Table regions from a PP-DocLayout_plus-L + churro-3B OCR run over 9.16M pages
            of the Deutsches Zeitungsportal}
}