Public Document AI Datasets: Which Ones Allow Commercial Use?

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Two layers of rights: annotations versus the document images

A commercial clearance needs two answers: who may license the annotations, and who holds rights in the pages those annotations sit on. Dataset authors usually own their bounding boxes, entity labels, question-answer pairs and class assignments, so the license on a card typically governs that layer. The page images are often third-party documents: tobacco-industry memos, scientific articles, annual reports, patents or government filings. A license from the annotators cannot grant more rights in those pages than the annotators held.

A license-compliance study of publicly available datasets makes this point directly: the license string on a dataset is the start of the analysis, and commercial-use rights depend on tracing every upstream source the data was built from [8]. For document AI, that tracing ends in an archive, a publisher or a crawl, not at the GitHub repository.

Per-dataset status table for common document AI benchmarks

The table below is a starting point for your own review, not a clearance: each row cites what a source in this session reported, and every row must be re-read on the official dataset page before you rely on it. Last checked: 9 October 2026.

Two patterns stand out. Three of the most-cited sets (RVL-CDIP, FUNSD, DocVQA) sit on the same tobacco-litigation archive, so one unresolved upstream question propagates across classification, extraction and question answering benchmarks. And the two "permissive" layout sets pair an open annotation license with page images drawn from many third parties.

Why the license string on a hub is not enough

Hosting-site license fields are frequently wrong or missing, so verify at the original source. The Data Provenance Initiative audit of more than 1,800 text datasets found license omission above 70% and license error rates above 50% on popular hosting sites [10]. DocLayNet illustrates the need to check the primary source directly: its GitHub repository LICENSE file states CDLA-Permissive-1.0 [5], a term not reliably surfaced on every dataset-hub mirror.

Document-specific surveys point the same way. The BigDocs authors report surveying 133 document datasets and finding roughly 80% with non-permissive or unclear licenses, leaving 16 they judged fully accessible and permissive [9]. If your team is building a commercial mix from public sets, expect most candidates to fall out at the license step.

Mirrors add a further failure mode. Re-uploads of RVL-CDIP or DocLayNet under personal accounts may carry a different license tag, a converted format (Parquet, COCO JSON) and no link back to the original terms. Pin the canonical source and record its URL and commit hash in your data card.

What "permissive" covers under CDLA-Permissive-1.0

CDLA-Permissive is a data license from the Linux Foundation's Community Data License Agreement family, written for datasets rather than code [11]. Its FAQ explains how the agreements treat results produced by computational use of data, which is the clause most relevant to model training [11]. Read the version that applies: DocLayNet's repository names version 1.0, while the Linux Foundation has since published 2.0.

What the license does not do is warrant third-party rights in the content. When annotators license their labels under CDLA, a downstream user still has to decide whether training on the underlying annual report or patent page is permissible in the target jurisdiction. That analysis sits outside the license text.

Commercial evaluation usually counts as commercial use

Using a non-commercial benchmark to score a product model is still a commercial use under most readings of "non-commercial, research and educational purposes" terms. FUNSD's reported terms [3] do not carve out evaluation. Teams often clear training data carefully and then run release-gating evals on restricted benchmarks without review.

The practical answer is a held-out set you are licensed to use for product decisions. The comparison on private evaluation sets versus public benchmarks covers the tradeoffs, and the general method for reading eval terms is on the public benchmark license page.

Benchmark fitness, not just rights

Rights are not the only reason to look past the classic sets. Much published document AI work reports results on a small set of public benchmarks such as DocLayNet and FUNSD, and RVL-CDIP's single-page, 16-class framing does not reflect production document streams. RVL-CDIP pages are grayscale scans from decades-old corporate archives, while production intake mixes born-digital PDFs, phone photos and multi-page packets.

If your target is invoices, bank statements, bills of lading or loss runs, the public sets mostly measure transfer from tobacco-era memos and scientific layouts. See document classification training data and layout analysis datasets beyond DocLayNet and PubLayNet for task-specific gaps.

Regulatory documentation: what a GPAI provider records

Providers of general-purpose AI models placed on the EU market owe Article 53 duties. The Code of Practice copyright chapter asks signatories to maintain a copyright policy covering those models [12], and the Commission's 24 July 2025 template sets the baseline for a public summary of training content [13]. A per-dataset clearance record like the one below feeds both.

Illustrative example: invented to show structure; it does not describe an available dataset.

dataset_clearance_record:
  dataset: DocLayNet
  canonical_source: <official repository URL>
  version_or_commit: <hash or release tag>
  last_checked: 2026-10-09
  annotation_license: CDLA-Permissive-1.0   # as named in the repository LICENSE file
  metadata_license_tag: <confirm per source>  # not all mirrors surface a license tag
  underlying_content:
    origin: mixed third-party pages (reports, manuals, patents, regulations)
    rights_basis: <counsel conclusion per jurisdiction>
  intended_uses: [pretraining, fine_tuning, internal_eval]
  release_gating_eval_allowed: <yes/no, with reason>
  attribution_required: <text and placement>
  pii_review: <method and sample size>
  reviewer: <name, role>
  decision: approved | approved_with_conditions | rejected

Clearance checklist for a public document dataset

Run every candidate through the same questions before it enters a training or eval mix:

1. Locate the canonical source (author repository or paper page), not a mirror, and record URL plus version.
2. Read the annotation license text in full, not the hub tag.
3. Identify the origin of the page images: archive, publisher, crawl or synthetic generator.
4. Find the rights terms for that origin, such as an archive copyright page or article-level licenses.
5. Check whether the dataset is a derivative of another set (FUNSD from RVL-CDIP, RVL-CDIP from IIT-CDIP) and clear each parent.
6. Decide separately for pre-training, fine-tuning, internal evaluation and release-gating evaluation.
7. Scan pages for personal data: names, signatures, addresses and account numbers appear on real forms.
8. Record the decision and a re-check date, because terms and cards change.

Pair this with SourceX's AI training data due diligence checklist for the broader vendor and rights questions, and with licensed vs synthetic vs scraped data when deciding what replaces a set that fails.

When public sets fail: options for commercial document models

If the clearance fails, the realistic routes are synthetic documents, licensed PDF corpora, or licensed operational documents from businesses that hold them. Synthetic generation avoids third-party page rights but has known distribution gaps; see where generated documents break. For broad multimodal pre-training, see licensed PDF corpora.

For extraction targets tied to real workflows, licensed business documents carry a rights basis you can document. SourceX sources operational datasets from US companies on request, including documents and finance and legal workflows; each dataset is rights-reviewed for ownership and consents and delivered under a license that defines records, uses, term and delivery. Buyers can describe the documents they need; a request does not guarantee a match. The document AI data hub maps tasks to annotation layers.

Licensed business documents for commercial document AI

SourceX finds US businesses that hold the documents you describe, and every release is approved by the supplying company. The process runs Find, Assess, Agree, Transact, Manage, and nothing is contracted until a supplier agrees. Describe your document data need to SourceX.