RfKnowledge/wp3-docduck-ocr-evaluation

WP3 Docduck OCR evaluation inputs

Private evaluation inputs for the WP3 multilingual OCR study, across 31 languages. The collection contains page images, the PDFs submitted to the OCR pipelines, generated reference text, text-block coordinates and the saved sample selections. It contains inputs and references, not model predictions or human-certified transcriptions.

Archive	Contents
full-pages.tar.gz	3,100 original canonical PNG pages, 31 original 100-page image-only PDFs, references, selection and generation settings. 620 development pages and 2,480 held-out pages.
pilot-pdfs.tar.gz	The 31 original five-page PDFs and 155 reference records used for the initial pilot. These pages are already included in development; they are not additional samples.
comparison-crops.tar.gz	496 original comparison-crop images recovered from the saved cluster preparation, 31 original 16-page image-only PDFs and their references and selection.
calibration-crops-reconstructed.tar.gz	620 reconstructed calibration crops from 310 original development-page regions, each clean/degraded, plus the two recorded selections. See the reconstruction note below.

Pages and splits

Every language has 100 pages: physical pages 1–20 are development, and 21–100 are held out. The initial pilot is development pages 1–5. The earlier full-page Gemini calibration uses pages 1–20 of Finnish, Irish and Maltese; those 60 pages are also already included.

Five layouts are represented: single column, two columns, three columns, prose table and sidebar, with sans/serif fonts. By page count, 60% use sampled source-sentence fragments, 20% use word-level Markov text and 20% use literal character controls. Development uses FLORES dev; held-out uses devtest, excluding articles whose URLs occur in development. Pages can reuse text and share templates; they are not all independent observations.

Canonical full-page images are 2,550 × 3,301 pixels, rendered at 300 DPI with CropBox. The PDFs contain images, without an OCR text layer. Generated references retain case, diacritics and the deliberately unusual spellings in controls.

Crops

The 496 comparison crops are a separate targeted development panel: four text blocks per language × two fonts × clean/degraded conditions. Two blocks are intact natural passages and two are literal controls. Thus the 248 natural images represent 62 distinct passages, two per language. These are the inputs shared by LightOn OCR 2, LightOn OCR 3 and the three Docling configurations, and by the targeted Gemini calibration. They are not 496 pages sampled from the full-page set.

The earlier calibration uses one declared region from each development page 1–10, across all languages: 310 regions × clean/degraded = 620 images. Its two selection YAMLs contain the original 30 pilot crops and the 590 added crops, with no overlap. The selected block's text in full-pages/candidates.jsonl is the reference. The filename includes its language, one-based page, block label and image variant.

Reconstruction note: the earlier calibration PNGs were no longer available in local temporary storage on 21 October 2026. This archive regenerates them from the retained canonical page PNGs using the saved crop_image procedure and Pillow 12.3.0. It uses the recorded boxes, scale 0.65, Gaussian blur radius 0.6 and JPEG quality 60, followed by resizing to the original crop dimensions. The reconstructed images have not been compared byte-for-byte or pixel-for-pixel with the original submitted calibration images. They are explicitly separate from the retained original 496-crop comparison panel. No new model inference was performed.

Reading the files

Download requires access to this private dataset and Hugging Face authentication. Extract the archives into the same directory. Each has its own top-level folder.

import json
from pathlib import Path

root = Path("extracted")
with (root / "full-pages/candidates.jsonl").open() as stream:
    pages = [json.loads(line) for line in stream]

development = [p for p in pages if p["split"] == "development"]
heldout = [p for p in pages if p["split"] == "heldout"]
image = root / development[0]["image"]
reference = development[0]["candidate_text"]

document_id is the language identifier; page_index is zero based. blocks contains the printed strings and their boxes; candidate_text is their newline-joined text. Image/PDF paths in the packaged reference and selection files are relative to the common extraction directory. Only those paths were relocated; reference strings, split labels, coordinates and image/PDF bytes are preserved for the original page, pilot and comparison inputs.

All 47,120 full-page block strings agree with their saved page reference assembly. The retained reference status remains unreviewed_synthetic: these checks do not certify every glyph, establish a uniquely correct reading order or demonstrate real-document OCR accuracy. Both LightOn models have processed the held-out pages, but held-out quality remained unscored when this collection was uploaded. Uploading does not change the split or authorize tuning on held-out results.

Source attribution

This is a custom WP3 experiment using Docduck layout helpers, rather than an unchanged upstream benchmark. Source text derives from FLORES-200, distributed under CC BY-SA 4.0 according to its dataset card. Changes include selection of sentences/fragments, Markov recombination and separate artificial controls.

Character inventories come from Unicode CLDR release 48. Noto Sans/Serif and their Georgian counterparts supply the fonts. The original Docduck, CLDR and Noto notices are included as LICENSE-Docduck.txt, LICENSE-CLDR.txt and LICENSE-Noto.txt. These notices describe their respective components; they do not relicense the source passages.