Document datasets for AI training. Real layouts, personal data masked.
Contracts, filings, invoices, forms, reports, and spreadsheets, licensed for training and labeled by the professionals who read them for a living. Every document is scanned for personal data and masked before it leaves our environment, and every item ships with a signed provenance record.
What is a document dataset?
A document dataset is a set of business documents, real or written to be indistinguishable from real, paired with the labels a model needs to learn from them: what kind of document it is, which fields matter, where the tables start and stop, which clause changes the risk, and what a reasonable question about the page would be answered with. The layout is part of the data. A model that has only seen clean text has never met a scanned form, a stamped filing, or a spreadsheet with merged header cells.
That is the gap most document-understanding projects run into. The material that would teach the model is locked inside firms, full of personal information, and rarely available with clear rights. Scraped PDFs and synthetic text either carry legal exposure or teach a tidier world than the one the model will work in.
eeko's data practice closes that gap in two ways. We license documents and tabular data from firms and practitioners who have signed off on AI-training use, and we commission realistic documents from the experts in our network where originals cannot leave the building. Both routes run through the same pipeline: a personal-data scan and mask at intake, labeling by people qualified in the field, second-expert review, and a signed provenance record on every item. It is the same material we rely on when we build intelligent document processing and retrieval systems for regulated clients, and it is part of the wider eeko AI training data practice.
Document datasets built to your schema.
Every engagement starts from a written specification and a pilot batch. These are the deliverables buyers combine most often.
- Consent and license terms bound per item
- Personal data masked, metadata stripped
- Near-duplicate templates caught at intake
- Authored by credentialed professionals
- Follows real structure and terminology
- Authorship recorded on every item
- Column-level documentation
- Reconciliation targets where relevant
- Masked identifiers, preserved structure
- Classification and key-value extraction
- Table cells, headers, and spans
- Label source recorded per field
- Grounded answers with citations
- Multi-page and cross-reference questions
- Summaries at agreed lengths
- Scan at intake, mask before delivery
- Metadata stripped from every file
- Masking decisions logged per item
What happens to a document between intake and your receipt.
An illustrative batch of commercial leases for a clause-extraction model, every step written to the record.
Expert labels on documents, from people who sign them for a living.
Layout tells you where a clause is, not whether it is material. That judgment comes from the field, so the labeling does too.
Matched by field, checked by a peer
- Attorneys and paralegals — clause classification, defined-term linking, obligation and liability tagging across agreements, filings, and correspondence.
- Accountants and financial analysts — line-item extraction, reconciling entries, statement structure, and the figures that tie across schedules.
- Insurance specialists — policy wording, endorsements, claims forms, and loss runs, labeled with the vocabulary adjusters and underwriters use.
- Clinicians and health administrators — intake forms, referral letters, and billing documents, with the fields that drive downstream decisions marked.
- Data analysts — schema documentation, column semantics, and the quiet errors that live in exported spreadsheets.
Experts clear a qualification round drawn from your guidelines before touching production data. Every label records who made it, whether it started as an AI suggestion, and what the reviewer decided. The same bench is available for annotation on documents you already hold.
From a one-page brief to a signed receipt.
Fixed scope and fixed price, agreed after the pilot. No volume commitment until you have seen a batch you would train on.
Document dataset questions.
What is a document dataset for AI training?
A document dataset is a collection of real or realistic business documents, such as contracts, invoices, forms, filings, and spreadsheets, paired with the labels a model learns from: classifications, extracted fields, table structure, layout regions, and question-answer pairs. Unlike a plain text corpus, a document dataset keeps the layout, so a model learns to handle a stamped scan or a spreadsheet with merged headers the way they show up in practice. eeko builds document datasets from licensed originals and expert-written documents, with personal data masked and a provenance record on every item.
How do you remove personal data from document datasets?
Every document and table is scanned when it enters our pipeline. Names, account and policy numbers, addresses, phone numbers, email addresses, and similar fields are detected and masked before anything is delivered, and embedded metadata such as author and location fields is stripped from the file. The masking log travels with each item. Masking lowers your exposure but does not replace your own privacy assessment, and your counsel should confirm the obligations that apply to how you use the data.
Are the documents real or written for the dataset?
Both, and each item says which. Licensed originals come from practitioners and firms that have signed a training license, and they ship with personal data masked. When the real file cannot leave the firm, as with live client contracts, attorneys, accountants, and other specialists write documents that follow the real conventions, structures, and quirks of their field. The origin is recorded on every item, so you can filter, weight, or exclude either category when you train.
Which labels can you add to documents and tables?
Document classification, key-value extraction, entity and relation tags, table cells and headers, layout regions, summaries, and question-answer pairs that cite the page they come from. Column-level descriptions and schema notes accompany tabular data. Labels follow a schema agreed with you in writing, are produced by specialists in the relevant field, and are checked by a second expert on an agreed share. Each label records who produced it and whether it started as an AI suggestion.
What formats do document datasets ship in?
Source files arrive in their native format, typically PDF, Word, or a spreadsheet format, with labels and metadata as JSON Lines, CSV, or Parquet. Tabular datasets can ship directly as Parquet or CSV with a written schema. Each delivery includes Croissant metadata describing the set and a signed receipt that lists every item by identifier and hash, so the exact contents can be reconstructed and checked later.
How much does a document dataset cost?
Price depends on the document type and length, whether items are licensed originals or commissioned, how deep the labeling goes, which expertise it requires, how much second-expert review you want, and whether you need exclusivity. A set of long commercial agreements with attorney clause labels costs more per item than invoice field extraction. Quotes are per document or per page, issued after scoping, and the pilot batch is priced the same way, so the number does not move once you are in production.
Related data solutions and expert fields.
Tell us which documents your model keeps misreading.
Send the document classes, the labels you need, and the volume. A data lead replies inside one business day with a sourcing plan and a pilot proposal. No deck. A specification and a price.
Experienced within
Markets served.
As an enterprise AI agency, eeko systems delivers production AI systems remote-first across the United States and internationally — including these markets: