DocuBench

A public benchmark for schema-guided structured extraction from 50 hard, real-world documents.

Each example contains a source document, a JSON Schema, and a hand-verified JSON label. The task is to extract the labeled structured data from the document according to the schema. Scoring is macro-average field accuracy with order-independent array matching, computed by an open standalone scorer.

Canonical repo, scorer, and raw baseline outputs: https://github.com/DocuPipe/DocuBench

Leaderboard

Committed baselines, scored by the public scorer against the hand-verified labels:

Rank	System	Accuracy
1	DocuPipe (high effort)	97.56%
2	DocuPipe (standard effort)	96.31%
3	Gemini	95.80%
4	GPT	93.54%
5	Extend	91.11%
6	Claude	90.33%

DocuPipe built this benchmark. Every system, including DocuPipe, is scored by the same open scorer against the same labels, with identical schemas, and every raw model output is committed under results/ in the GitHub repo so the table is reproducible.

Composition

• 50 documents, 50 schemas, 50 hand-verified labels
• 10 file types: PDF, JPEG, PNG, TIFF, XLSX, CSV, XML, TXT, DOCX, HTML
• 11 languages/scripts: English, Hebrew, Japanese, Chinese, Arabic, French, German, Portuguese, Dutch, Italian, Spanish
• Failure modes covered: line-item arrays, multi-page tables, reconciling totals, right-to-left and CJK scripts, rotated scans, handwriting, nested objects, needle-in-haystack lookups

Documents cover invoices, bank and brokerage statements, utility bills, annual reports, payslips, purchase orders, waybills, lab reports, discharge summaries, engineering drawings, insurance declarations, tax forms, spreadsheets, XML, CSV, text, and HTML.

Motivation

Many document extraction evaluations focus on single-page, flat, or QA-style tasks. DocuBench focuses on end-to-end structured extraction into realistic JSON shapes, including arrays, nested objects, multipage context, non-Latin scripts, and non-PDF inputs.

Collection and labeling

Documents were selected from public sources, vendor sample documents, government publications, open datasets, and benchmark-authored synthetic files. Each document has a source and license record in SOURCES.md and sources.json in the GitHub repo. Labels were authored for the benchmark and manually checked field by field against the source document.

Intended uses

• Evaluating document extraction systems
• Testing schema-guided extraction robustness
• Comparing parser or extraction workflows on public artifacts
• Regression testing extraction systems across file types and languages

Out-of-scope uses

• Training models on the test labels
• Claiming broad document AI superiority from the headline aggregate alone
• Evaluating privacy handling, security, or compliance controls
• Treating these 50 documents as representative of all enterprise documents

Licensing

• Code is MIT licensed.
• Labels, schemas, benchmark-authored metadata, and benchmark-authored results are CC BY 4.0 unless stated otherwise.
• Source documents retain their original licenses or publication basis.

Citation

If you use DocuBench, cite the GitHub repository: https://github.com/DocuPipe/DocuBench