biglam/britannica — page store

Page images of the Encyclopaedia Britannica, 1768–1929, from Internet Archive scans: 1,160 volumes, ~977,000 pages, full resolution. Public. This bucket is the image store behind the dataset biglam/britannica-illustrated-pages; the raw IA items it was built from are in biglam/britannica-mirror.

Layout

source/pages/<ia_id>/<member>.jp2      originals, one file per page, IA member names kept verbatim (~1.08 TB)
source/jpg/<ia_id>/<member>.jpg        2000 px JPEG copies of the pages in the illustrated-pages dataset (~115k files)
predictions/<model>/part-*.parquet      model outputs over the pages, one directory per model
manifest/                               build and verification logs (see below)

<member> is the file name inside the IA _jp2.zip for the volume, e.g. encyclopdiabri20chis_0784.jp2. The trailing number is the IA leaf, so https://archive.org/details/<ia_id>/page/n<leaf> opens the same page on archive.org. Some IA items hold several volumes; their members are prefixed with the volume name (Volume 1, A-AME_0784.jp2).

Fetch a page:

https://huggingface.co/buckets/biglam/britannica/resolve/source/pages/<ia_id>/<member>

From a dataset row to a page

Rows of biglam/britannica-illustrated-pages carry bucket_url (the original), bucket_url_jpg (the JPEG copy) and leaf. Every row was matched to its page by comparing the dataset's embedded thumbnail with the store image (grey-level correlation ≥ 0.99) before the links were written; manifest/rowmap-*.jsonl records the match for every row (file_key, member, leaf, corr, width, height).

Predictions

predictions/crop-masks-seg-v2/ — instance boxes and masks for illustrations on every page of the illustrated-pages dataset, from small-models-for-glam/britannica-illustration-detector-seg-v2. One row per page: objects (bbox [x, y, w, h], score, area, rectangularity) and masks_rle (COCO RLE), all in the original page's frame (src_width × src_height). Stored down to score 0.10; 0.3 is a sensible display cut. Masks are the raw model output; for cut-outs, dilate by ~2 % of the box size. The same data is published as the crop_masks config of the dataset.

Provenance

• manifest/explode-shard-*.jsonl — per-volume record of the extraction from the IA archives (pages, bytes, seconds).
• manifest/rowmap-shard-*.jsonl — per-row page match for the dataset (above).
• Mirror: biglam/britannica-mirror, _manifest/catalog.parquet lists the IA items and files.

Scans are public domain. Built by Daniel van Strien with Hugging Face Jobs and bucketbag, 2026-08.