niles-liu/strikethrough-benchmark

pdf-strikethrough-detect benchmark assets

Files the pdf-strikethrough-detect benchmarks read or wrote that are not kept in the code repository: scanned-recovery/ for the package's own benchmarks, and frontier/ for its benchmark against frontier models.

scanned-recovery/

Azure Document Intelligence results for the pages benchmarks/scanned_recovery.py scores, which produce the Azure DI figures in the package's "Choosing an OCR backend" table. Each file is the analyze-result JSON (prebuilt-layout, API version 2024-11-30, with its analyzeResult envelope) for an image-only PDF of the listed pages of a public redline document, rendered at 200 dpi. With them, the benchmark scores Azure DI's words without an Azure account. benchmarks/fetch_corpus.py downloads them and checks each against the sha256 in benchmarks/manifest.json, so a changed file fails loudly instead of quietly moving a number.

File	Source document	Pages (0-based)
copyright-201-10-redline.scanned.di.json	US Copyright Office, 37 CFR 201.10 proposed rule (redline)	0–3
fdic-fair-lending-redline.scanned.di.json	FDIC, Fair Lending laws & regulations (redlined changes)	0–2, 23–25, 33, 34, 37, 42, 43, 53
gretna-udo-edits.scanned.di.json	Gretna, LA, Unified Development Code substantive edits (redline)	2, 3, 11, 14, 18–21

The source PDFs are public documents published by those bodies; benchmarks/manifest.json lists each one's URL and sha256. To reproduce the figures:

git clone https://github.com/niles-liu/pdf-strikethrough-detect
cd pdf-strikethrough-detect
pip install -e ".[rapidocr]"
python benchmarks/fetch_corpus.py
cd benchmarks && python scanned_recovery.py

frontier/

The data of a pre-registered benchmark of ten frontier models against pdf-strikethrough-detect 0.12.1 at finding struck-through text. Its results, rules, reasons and deviations are in the package's benchmarks/frontier/: start with RESULTS.md. One folder per run, as the harness wrote it:

Path	Holds
<run>/requests.jsonl	every request: its arm, page, model, rung, input and settings, and the sha256 of its prompt and of its input
<run>/responses.jsonl	every attempt: its status, text, usage, the model the provider reported, timings and cost, and the commit, package and tags it ran under
<run>/scores.jsonl, <run>/provenance.json	per arm and page, the counts behind every metric; the commit and tags it was scored at, and the files changed since prereg-2
<run>/*.csv, <run>/*.json, <run>/*.txt	the run's report tables
free/	the free arms (the package and two baselines) on every page, and their timings
test/gallery.html, test/scatter.svg	each arm's worst failures on the test pages; leakage against cost
pilot/audit/, g2-audit/, g1-audit/, test/audit/, test/g1-relabel/	the hand audits: the page each was labelled on (A.html), its key (key.jsonl), the labels as given (A.csv), any corrected or reviewed copy with a note saying why, and the labels as read in
g0/dev/, g0/test/, g0/G0-qa.jsonl	G0's synthetic pages, as PDFs, and the questions the QA probe drew from, each with its live and struck answer
rapidocr/	the three OCR model files bundled with RapidOCR 3.10.0, which the package's scan path ran, each matching the sha256 the free arms recorded, with RapidOCR's LICENSE and MODEL_LICENSES.md
manifest.json	each file's sha256 and size

The runs: smoke, smoke-batch and cache-probe checked each provider; pilot, ladder, ladder-tiled, paraphrases, repeats, ablation and flagship-check ran on dev pages to choose the prompts and settings; canary-1 to canary-4, test, fable, qa and latency are the test run. test/g1-relabel/G1-corrections.jsonl holds the relabel's corrections to G1's ground truth.

A page ID names its set: G0, synthetic redlines generated from public seeds; G1, pages of the public redline documents in the package's corpus_manifest.json; G2, US Senate reported bills from GovInfo; G5, handwritten pages from BRESSAY. The page images themselves are not here; the audit pages and the gallery hold crops of them. The HTML pages (test/gallery.html and each A.html) work only once downloaded: the Hub serves them sandboxed, without their scripts, styles or images.

The prompts are in the package's benchmarks/frontier/prompts/. Left out: xAI's team ID in its error texts and the Google Cloud project number in Vertex batch IDs, each replaced by [scrubbed]; run logs and batch bookkeeping; the harness's code, which is to follow; and the held-back G0 seed, whose sha256 alone is published.

Cite the benchmark by its Zenodo record, 10.5281/zenodo.23425895.

License

scanned-recovery/: CC0 1.0, to the extent the uploader holds any rights. The files are machine OCR of the source documents: the Copyright Office and FDIC documents are US federal government works, not subject to US copyright (17 U.S.C. §105); the Gretna document is a City of Gretna, Louisiana public document whose text stays under its publisher's terms.

frontier/: the labels, scores, tables and notes are CC0 1.0. The model responses (responses.jsonl, and the outputs quoted in test/gallery.html and the audit pages) are not covered by that grant: they are shared so the results can be reproduced and audited, each provider's terms govern them, and several forbid using model output to train or develop AI models, so do not use them for that. Text and images that reproduce a page keep its source's terms:

• G0's pages are generated by the benchmark: CC0 1.0.
• rapidocr/ mirrors RapidOCR's files unchanged, under the Apache License 2.0: see its LICENSE and MODEL_LICENSES.md there.
• G2's bills and G1's federal documents are US government works, not subject to US copyright (17 U.S.C. §105); G1's California documents and the Gretna, Louisiana code stay under their publishers' terms.
• G5's pages are from BRESSAY (10.5281/zenodo.11637681), licensed CC BY 4.0. Here are excerpts of its page transcriptions in test/gallery.html, marked up (struck text struck through, the failing run highlighted) and cut, and the systems' readings of its page images in the responses; no BRESSAY image is included. Its authors ask that it be used for non-commercial research and teaching only, and that published work cite: Neto, A.F.S., Bezerra, B.L.D., Araujo, S.S., Souza, W.M.A.S., Alves, K.F., Oliveira, M.F., Lins, S.V.S., Hazin, H.J.F., Rocha, P.H.V., Toselli, A.H.: BRESSAY: A Brazilian Portuguese Dataset for Offline Handwritten Text Recognition. In: 18th International Conference on Document Analysis and Recognition (ICDAR). Springer (2024).