You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

scanned_docs_internet_archive_part2

Raw scanned and photographed document page images collected by the ElliotVL DataEngine as an OCR / document-understanding image pool. 5,006,279 images, 3.96 TB.

This repo holds part of a source that is split by shard range across several repos (the Hub caps public storage per org). The other parts: VL-Data/scanned_docs_internet_archive_part1.

  • Internet Archive β€” page images unpacked from the per-item page-image archives (_jp2.zip / _images.zip / _tif.zip) of modern-language collections, one config per language. Source: https://archive.org/help/aboutsearch.htm
config source images shards size
spanish ia_spanish_modern 736,342 80-434 of 435 0.38 TB
swahili ia_swahili_modern 8,246 0-2 of 3 0.00 TB
swedish ia_swedish_modern 362,332 0-164 of 165 0.18 TB
tagalog ia_tagalog_modern 10,957 0-3 of 4 0.00 TB
tamil ia_tamil_modern 425,975 0-101 of 102 0.11 TB
telugu ia_telugu_modern 938,854 0-498 of 499 0.54 TB
thai ia_thai_modern 45,623 0-22 of 23 0.02 TB
turkish ia_turkish_modern 36,983 0-26 of 27 0.03 TB
ukrainian ia_ukrainian_modern 173,854 0-196 of 197 0.21 TB
urdu ia_urdu_modern 1,999,154 0-854 of 855 0.92 TB
vietnamese ia_vietnamese_modern 47,646 0-17 of 18 0.02 TB
welsh ia_welsh_modern 23,517 0-11 of 12 0.01 TB
yiddish ia_yiddish_modern 116,838 0-64 of 65 0.07 TB
irish ia_irish_modern 79,958 0-1383 of 1384 1.47 TB

What this is, and what it is not

  • Unlabeled. There is no OCR text, caption or QA. Each row is one page image plus its provenance; labels are meant to be produced downstream.
  • Original bytes. Every image is stored exactly as downloaded (JPEG, JPEG 2000, TIFF, GIF), not resized or re-encoded. Pages are large: census frames are typically ~7096Γ—5472, Internet Archive pages ~4000Γ—6000. Decode JPEG 2000 with a Pillow built against OpenJPEG.
  • Deduplication is exact by downloaded file / zip member within a source. image_sha256 is the hash of the stored bytes, so exact duplicates across sources can be removed by joining on it. Near-duplicates (the same page scanned twice) are not removed.

Schema

  • image β€” HF Image(); path is the original file name
  • image_sha256 β€” SHA-256 of the stored image bytes
  • source_id β€” collection lane, e.g. nara_catalog, loc_maps, ia_german_modern
  • item_id / item_url β€” the record or book the page belongs to (NARA catalog id, LOC item, Internet Archive identifier)
  • image_url β€” the URL the image was downloaded from (empty for pages taken out of an IA archive)
  • title β€” record / book title when the source provides one
  • page_index β€” 0-based page position inside an Internet Archive item, else empty
  • width, height, image_format, file_name

Parquet shards are ~1 GB with ~32 MB row groups, 1000 shards per subfolder.

How this was made

Code and design notes: Elliot-Data/elliotvl-data-pipeline. Start with DataEngine/docs/scanned_doc_pipeline.md (download lanes, resumable state, IA page manifests), the run log DataEngine/reports/scanned_doc_download_run_20260702.md, and for this Hub layout DataEngine/scripts/pool_publish/scanned_docs_to_hub.py / scanned_docs_publish.py.

Rights

Internet Archive. Items were uploaded to the Internet Archive by many different contributors and their rights status varies. Many of these books may still be under copyright in some jurisdictions. This copy is for research use; you are responsible for checking each item at item_url before any other use.

Access is gated with manual approval.

Downloads last month
264