illustrated_ads

# 19th Century US Newspaper Adverts — illustrated or text-only 549 advertisement images cut from digitised United States newspaper pages in the Library of Congress [Chronicling America](https://chroniclingamerica.loc.gov/) collection, each labelled `text-only` or `illustrations`. The adverts were located by [Newspaper Navigator](https://news-navigator.labs.loc.gov/) (LC Labs), which ran an object detection model over 16,358,041 Chronicling America pages to extract visual content; advertisements are one of its categories. The sample is even by year, not by title: 61 adverts from each of nine years at five-year intervals, 1860–1900, drawn from 301 newspaper titles across 245 places of publication. Sampling and annotation used [nnanno](https://github.com/Living-with-machines/nnanno). It was built as teaching data for the Programming Historian lesson *Computer Vision for the Humanities* ([10.46430/phen0101](https://doi.org/10.46430/phen0101)) — a realistic rather than clean example for teaching image classification on digitised heritage material. ## Labels | label | count | |---|---| | `text-only` | 376 | | `illustrations` | 173 | Roughly 2:1, so accuracy is a weak metric here — report per-class scores. ## Structure One `train` split, no train/test division. Each row carries the cropped advert (`image`), the `label`, OCR text for the advert region (`ocr`), the detection box on the source page as relative coordinates (`box`) and its confidence (`score`), plus Chronicling America metadata: `pub_date`, `lccn`, `name`, `publisher`, `place_of_publication`, `geographic_coverage`, `batch`, `page_seq_num`, `edition_seq_num`, `file`, and `url` / `page_url` pointing back to the source crop and full page scan. > [!WARNING] > The images are single-channel greyscale (PIL mode `L`). Call `.convert("RGB")` before feeding them to a pretrained backbone. Two further traps. Every row has `score` ≥ 0.90 (range 0.900–0.998), so adverts the detector found ambiguous never entered the sample — this data is easier than the page collection it came from. And `ocr` is raw Chronicling America OCR of the crop: noisy throughout, and empty for 21 of the 549 rows. ## Uses and limits Binary image classification, transfer-learning walkthroughs, and a small test bed for how the illustrated/text-only split changes across 1860–1900. At 549 rows it is too small to support claims about the 19th-century press: the 301 titles here are the ones Chronicling America has digitised, and that programme selects title by title through state partners. ## Licence and citation Public domain (CC0-1.0), per the [Zenodo deposit](https://doi.org/10.5281/zenodo.5838410); the underlying page scans are US public domain material held by the Library of Congress. The Zenodo record credits Daniel van Strien (British Library) as author and Catherine Bond-Harris as data collector. ```python from datasets import load_dataset ds = load_dataset("biglam/illustrated_ads", split="train") # 549 rows, ~48 MB image = ds[0]["image"].convert("RGB") ``` ```bibtex @dataset{van_strien_daniel_2021_5838410, author = {van Strien, Daniel}, title = {{19th Century United States Newspaper Advert images with 'illustrated' or 'non illustrated' labels}}, month = oct, year = 2021, publisher = {Zenodo}, version = {0.0.1}, doi = {10.5281/zenodo.5838410}, url = {https://doi.org/10.5281/zenodo.5838410}} ```