Best document and OCR dataset providers in 2026

1. Does the sample cover the documents your model will see?

An invoice, a handwritten form and a long report pose different collection problems. Start with document families, languages, layouts and scan conditions. Then settle the counting unit. Ten pages from one report give you a different kind of variety from ten unrelated forms.

Ask for counts by document family and source or template group, plus examples from the difficult groups. For multi-page work, inspect complete documents using the page-order review. An owner can prepare the same breakdown before approaching a licensing partner. A single archive-size total leaves too much of the eventual offer undefined.

Call test: “Show us each required group, its counting unit and a complete sample. Which groups are unavailable?” Five attractive pages with no explanation of the larger collection are a reason to keep investigating.

2. What labels will arrive with each page?

Source images, machine-generated OCR and human-checked transcription are different deliverables. Field extraction adds another layer: which value belongs to which field, including blank, illegible and ambiguous values. Tables and reading order need their own definitions. For geometry, use the layout and coordinate review to inspect how labels connect to the page.

Call test: “Give us one complete document, its label files and the annotation instructions. Show a table, an empty field and unreadable text.” Load the pair into your pipeline. A promise of “annotated documents” without a schema cannot tell you what remains to build. When the missing work is a specified set of labels, a dedicated annotation service is a stronger starting point than an archive license alone.

3. Do the permissions cover the intended use and every included component?

Match the file list to the rights evidence. Ask who can authorize each source group and which uses the proposed agreement covers. Evaluation, training and redistribution are separate questions. A document containing photographs, client records or third-party text needs a scope that addresses those components.

Call test: “Map this sample to its source and licensing authority. What evidence can our counsel inspect, and what is excluded?” For buyer-supplied files, identify what evidence the buyer must provide. “We have the files” does not answer either question. Owners should make their rights inventory as specific as their document inventory before comparing proceeds.

4. Can you test the promised quality on a representative sample?

A correctly transcribed sentence can still have the wrong field label. A correctly labeled page can still be missing its continuation. Set separate checks for the outputs your task uses, then select a sample that exposes hard document groups. Ask how recurring templates and related documents are handled across any training and evaluation split.

Call test: “Show the sampling method, errors by label type and how disagreements were resolved. Which checks will apply to delivery?” A headline accuracy percentage with no denominator or task definition gives you no usable acceptance test. Keep difficult groups visible instead of letting their errors disappear into an overall average.

5. What does the complete offer include, and who pays for the remaining work?

Compare the same scope: rights, collection, digitization, annotation, redaction, review, delivery and any refresh you need. A page rate can be useful when those responsibilities are explicit. It becomes misleading when one quote includes reviewed labels and another leaves the buyer to commission them.

Call test: “Price this scope, including the minimum commitment, exclusions and payment milestones.” Owners should add: “Who pays each fee, what amount is my share calculated from, and when does payment become due?” Record preparation costs separately from licensing proceeds. A percentage without its calculation base cannot tell an owner what the offer is worth.

Published commercial information, checked November 2, 2026; LXT quality support rechecked November 3. “Not published” identifies an unanswered offer question, not a term the provider cannot negotiate. Re-check prices and terms before signing.

1. Rightmark: best for licensing publisher and expert archives

We help owners license existing material to AI teams while keeping ownership. Our document route covers publisher backlists, news archives, translation memories, courseware and forums. The useful starting point is an identifiable collection and its rights. Our buyer page specifies “Text by title or archive, with a rights record per item”; news files include dates, bylines and sections. That gives buyers a concrete delivery to inspect and owners a preparation list before discussing proceeds. For translation memories, language codes and domain metadata help define the collection. Scanned print archives can be digitized before delivery.

Training, retrieval and display are “Separate rights, priced separately.” Our terms require a separate signed licensing agreement. Buyers can inspect representative scrubbed samples under NDA, “for evaluation only”. Our fee routes state: “We take nothing from your payout on standard licenses: the buyer pays our fee on top.” Premium and bespoke deals leave owners 75% to 85%, agreed in writing before sale. Buyers cover extraction, scrubbing and transfer; owners pay nothing upfront. This makes the fee route visible before agreement: compare the calculation base, resulting owner proceeds and complete buyer total.

A text archive does not by itself supply labeled invoice images, word boxes or table structure. For that work, compare a dedicated OCR catalog or annotation service and specify the files and acceptance checks. Rightmark's public document offer does not establish a ready ground-truth catalog. Owners exploring this route can start with an indicative appraisal, then establish which works and uses belong in the agreement.

Choose Rightmark for an owner-specific publisher or expert archive; buyers needing prepared OCR labels should compare specialist annotation offers.

2. Shaip: best for commissioned OCR collection and annotation

Shaip, now part of Ubiquity, offers three routes: license an OCR dataset, commission document collection or commission annotation. Its OCR service covers invoices, receipts, handwriting and tables, with character, word and row transcription alongside spatial labels. That makes the first buying decision quite practical. If the source documents already fit, specify the labels you need. If they do not, write the missing document families and capture conditions into a collection brief before discussing annotation volume.

In Shaip's May 2026 OCR case, an unnamed client needed text positions and exact transcription across documents and real-world images with mixed scripts, distortion and poor legibility. The work paired word boxes with character transcription and tagged clarity, including partially legible and fully illegible text. Separate review checked the boxes and the transcription; unreadable text was flagged rather than guessed. The client, identified only as Director, Document AI, said: “Their dual QA on both bounding boxes and transcriptions gave us training data we could deploy.” The useful specification detail is that two outputs had two checks. Ask to see where those checks appear in the sample you receive, especially when a plausible transcription could conceal a missed or badly positioned word.

The catalog needs a closer look before a buyer selects it as ready ground truth. Several invoice and receipt image entries are marked unannotated; a handwritten-document entry is marked annotated. Obtain the exact image and label files for the offered lot. Pricing depends on dataset type, volume and complexity, with a custom quote. Shaip's customer terms make payments non-refundable unless the applicable agreement provides otherwise. Put required review and correction work into that agreement. For an archive owner, this quote is the cost of preparing specified files. Compare that cost with the scope of the license you intend to offer.

Choose Shaip for specified OCR collection or annotation; buyers choosing a catalog lot must check its label state before budgeting it as ready ground truth.

3. Label Your Data: best for OCR annotation of buyer-supplied documents

Label Your Data is a managed annotation service for teams with documents to prepare. Its OCR workflow normally starts with client-supplied data; collection can be requested with the type, volume and acquisition method specified. That distinction suits an AI team that already holds invoices or financial documents but still needs a usable training set. Before commissioning the work, identify the transformation you need: extraction, verification, labels or a combination. Buying more source pages will not resolve an undefined label specification.

The service agrees policies, labeling criteria and tools with the client, runs a small pilot, then moves to a dedicated team when the pilot is satisfactory. Production proceeds in batches with feedback, progress reports and QA. Its general workflow also supports the client's preferred annotation platform or its own. These are useful points to test together: return a difficult page with a disputed value, see how the instructions change, then inspect the next batch against the revised rule. In its published Hypatos testimonial, Luca van Skyhawk, identified there as VP of Customer Success at Hypatos GmbH, says: “Their flexibility and ability to work with multiple languages are impressive.” The surrounding account concerns training data for different document types and languages. The language mix in your project still belongs in the sample and agreed scope.

The OCR quote depends on document volume, complexity, languages and team training. Its pricing page offers an estimate from sample data, while the homepage lists hourly, per-object, per-task and project billing. Ask which basis applies and whether it includes review and rework. General prices for boxes or entities do not price a complete OCR project. Buyers missing the source documents need collection and permissions scoped separately. An archive owner can use the pilot to decide whether paying for preparation is justified by a proposed license. Agree who pays that bill before full annotation.

Choose Label Your Data to annotate documents you already hold; buyers without source material need a separately scoped collection and permissions plan.

4. Appen: best for configurable document annotation workflows

Appen's AI data service includes documents and OCR, while its ADAP platform supports text extraction, image annotation and configurable review stages. The fit is a document program that needs deliberate workflow design. Decide where a label is created, where it is checked and who resolves the difficult cases. A faint total on a receipt, for example, needs an instruction for uncertainty as well as a place to type a number. Put that decision into the task before scaling the work.

ADAP offers routing rules, internal-expert participation, contributor monitoring, gold test questions and validators. Appen also describes agreement thresholds set before work, adjudication of disagreements, and consent, licensing and lineage records for review. Ask which checks measure correctness for your field definitions. Two annotators choosing the same wrong value can agree, so agreement alone should not be your entire acceptance test. In Appen's account of general text and image annotation, Erica Nishimura, identified as Data Curator at GumGum, said she would “look to Appen to help me with the job design, creation, and coding.” That comment concerns job design; evaluate the proposed OCR workflow on your own representative documents.

A ready receipt corpus is a separate purchase decision. Appen's English US receipt listing remains in development, with QA and annotation still described as unfinished. Confirm completion and inspect the actual files before planning around catalog delivery. For custom work, scoping produces a per-unit rate and volume schedule; expertise, complexity, languages and the required agreement threshold affect pricing. Write the promised coverage and review work beside the quoted unit, then identify what triggers payment and correction. For owner-supplied files, write down which permissions the service needs and which uses remain outside the engagement. Keep those boundaries with the source records.

Choose Appen for a configurable document-data program; buyers needing a ready receipt corpus must resolve the listing's in-development status first.

5. LXT: best for custom document and OCR datasets

LXT offers custom OCR and document datasets, including annotation of a buyer's existing corpus. Its document offer describes analyzing the layout and quality of existing documents before proposing collection volume. A buyer can specify layout regions, key-value pairs and table structure. Its OCR offer separately describes word or line geometry with transcription, plus formats including ALTO XML and PAGE XML. Select the layers your model consumes, then load an image and its labels together before agreeing to the larger package. A supported format name is only the beginning of that integration check.

The text-collection workflow adds pilot calibration, gold tasks, automated checks and expert review, with metadata, task logs and documented QA. Request the records that explain a disputed label, along with the corrected output. LXT also publishes a historical video-project testimonial in which an unnamed Operations Lead at a leading AI-powered visual assessment company praised “the level of data quality and consistency that LXT delivered.” That video-collection comment concerns a different modality. For document work, inspect the promised transcription, fields or table relationships against the actual sample. Ask how a mistaken field boundary is corrected and how the same instruction reaches subsequent batches. LXT's collection offer, which includes OCR, commits to working with dissatisfied customers toward their expected quality. Agree document-specific acceptance checks and remedies in the statement of work.

LXT's price examples distinguish the work packages. The OCR page lists $10,000 for a focused single-language dataset of 5,000 to 20,000 text lines. The document page lists $12,000 for one document type with 1,000 to 5,000 pages. Both give $100,000 or more for larger multilingual work; the document example includes complex table and form annotation. These are project examples, not per-page rates, minimum commitments or a ceiling. Compare the included labels, collection and review before treating the two starting figures as alternatives. An owner commissioning this preparation still needs a separate licensing agreement if the aim is to earn proceeds from the archive.

Choose LXT for a custom document and label specification; buyers seeking immediate catalog delivery should confirm availability separately from commissioned work.

6. iMerit: best for reviewing document extractions and training labels

iMerit, now part of EXL, offers human review of document extraction, including validation and corrections. Teams can work in the buyer's annotation tool, iMerit's managed platform or third-party tools. That is a useful fit when documents and an extraction pipeline already exist, but the outputs need checking before becoming training labels. Its current text tool supports field and entity labels, centralized instructions, imported pre-labels and adjudication. An imported prediction should enter the review process with a clear status: proposed, corrected or accepted.

In iMerit's 2021 CrowdReason case, inaccurate document extractions left employees doing manual correction. Three reviewers assessed outputs; consensus-approved results became training data, while unresolved cases went to a team leader with source information. Brandon Van Volkenburgh, identified in the account as CrowdReason's CTO and co-founder, said the data “helped us get up and running with the development of our property tax software.” The practical question is who has authority to accept a label. Ask to follow one disputed field through that decision, with its source image and the rule used to resolve it. That exercise reveals more about a proposed review process than a demonstration containing only easy examples.

A buyer can use iMerit's configurable roles, routing and handoffs to discuss how labeling and review fit its own pipeline. Agree the schema, exception policy and responsibility for changing instructions before comparing the quote. The document service mentions service-level agreements but publishes no common document-annotation rate or universal acceptance remedy. Request a cost breakdown for annotation, review, adjudication and tooling. If the purchase also requires new source documents and permission to train, put those into a separate explicit scope. An owner checking extracted fields should retain the link from each accepted label to its source document. The eventual buyer can then inspect the correction in context.

Choose iMerit to review extractions and produce specified labels; buyers needing new source documents or a licensed corpus must scope those requirements separately.

7. CloudFactory: best for managed document annotation with client feedback

CloudFactory's Data Foundation offer combines collection, curation, dataset construction and annotation, including metadata enrichment and AI-assisted pre-labeling. For document teams, this is a way to discuss the work around the labels as well as the labeling itself. Decide which preparation steps you already perform and which the service will own. The current security page also distinguishes included network security from an endpoint upgrade. If sensitive files require the latter, include that requirement in the scope and cost discussion before anyone transfers a sample.

Heretik's legal documents varied in structure and wording, so its training labels needed to capture distinctions among agreements. In CloudFactory's published account, a managed team learned the task through training with client input and short feedback cycles, producing labels for contract structure, entities and clauses. Heretik co-founder Andy Abbott explained: “We wanted to work with a group to build domain knowledge so they could become more effective as time went on.” The buyer's contribution is part of that model. Identify the person who can explain why two similar clauses need different labels, and give that person a route for correcting instructions when the team encounters an unfamiliar document.

CloudFactory's pricing page offers free analysis of instructions and a small representative dataset before a proposal, expressly distinguishing it from a free trial. Use those returned examples to test your instructions. Consumption-based pricing carries annual spend commitments and customized rates; statements of work also allow piece-rate options and define service levels. Establish which arrangement you are being offered, how corrections are charged and how unused commitments are handled. A bounded evaluation does not itself cancel an annual commitment. This model suits buyers prepared to contribute expertise and feedback. If the team will need regular access to your subject specialist, reserve that person's time in the project budget as well as the service spend.

Choose CloudFactory for managed document labeling with active client feedback; compare the consumption plan's annual spend commitment before treating it as a small one-off purchase.

Choose two or three providers for the same work

Start with the work you still need: rights to an existing archive, new collection, annotation of your files or a specified combination. Narrow to two or three providers against the same document families, languages, labels, budget and commitment constraints. A service-only provider can be the right preparation partner without being a licensing partner.

Record dataset information and sample questions in the data card. Keep prices, payment milestones and contractual acceptance terms alongside it. The following scripts work across the shortlist. The red-flag answers are invented examples, not quotations from any provider.

Ask the same seven questions

1. “Which document groups can you deliver, in what quantities, and which still need collection?” Request counts by family, language, template or source group and scan condition. Inspect complete documents. Red flag: “A million pages; these five clean images tell you everything.”
2. “Show one complete document and every file we receive. Which text and labels are generated, reviewed or excluded?” Test the schema in your pipeline, including missing and illegible values. Red flag: “The PDFs are searchable, so all ground truth is included.”
3. “Who authorizes each source group, and what covers evaluation, training and later sharing?” Request the authority, exclusions and proposed terms for counsel to inspect. For your own files, establish what evidence you must supply. Red flag: “We have the files, so you can use everything however you like.”
4. “Who can access the files, what is removed, and what can be retained or reused afterward?” Name the party responsible for redaction, transfer, reviewer access and deletion. Put subcontractor access and permitted reuse into the handling plan. Red flag: “The NDA covers everything; access and retention need no detail.”
5. “Can we agree a representative sample and check results by document group and label type?” Define the checks, denominators and disagreement process. Ask how related documents and recurring templates affect any training/evaluation split. Red flag: “Accuracy is 99%, but we cannot explain what was counted.”
6. “What is the price unit, committed spend, included preparation and review, and every separately charged item?” Compare full costs and due dates. Owners should also request the payer and fee calculation base. Red flag: “Commit to the page rate now; we will price the required labeling later.”
7. “Which acceptance checks and payment milestones go into the agreement, and what happens if delivery fails?” Name the reviewer, review window, correction charges, recheck process and any agreed stop or refund option. Red flag: “Delivery means acceptance; we can discuss failed checks afterward.”

Agree on a bounded pilot before the full commitment

Propose 30 to 60 days of bounded work, subject to permitted sample access and the actual offer. For an existing corpus, that might mean authorized sample evaluation and a defined initial delivery. For collection or annotation, specify a limited batch. Write down the start condition, end date, budget, document groups, volume, schema and deliverables.

Agree acceptance before the work begins. Name the reviewer, checks and task-specific thresholds. Keep difficult groups visible in the results, record unreadable content and define which defects require correction or rejection. Set who pays for rework, when buyer charges and any owner payment become due, and what happens to sample access and retained files if the purchase stops.

End with a decision: accept and expand, correct and recheck, or stop under the agreed terms. Set the escalation contact and any extension charges beforehand. If a provider cannot offer this pilot, ask for a bounded evaluation alternative and its reason. An evaluation window creates neither an automatic refund nor a right to cancel a larger commitment.

Choose the supplier whose documents, labels, permissions and acceptance terms answer your written requirement. Resolve the gaps that would change that choice before committing to the full volume. For an archive owner, the corresponding decision is whether the proposed proceeds justify the preparation work and rights being licensed, with the payer and calculation base clear.

Before your next supplier call, complete the data card with the document coverage, labels, rights evidence and sample questions you still need answered. Keep price, payment and acceptance terms alongside it so you can compare complete offers.