Custom AI datasets. Sourced to the spec, not the scrape.
Tell us what your model keeps getting wrong. We write a dataset specification with you, fill it from consenting creators and credential-checked experts, check every item, and ship it under a licence written for the use you named. Pilot batch first, production at the same bar, fixed price on both.
What is a custom AI dataset?
A custom AI dataset is training data assembled against a written specification instead of pulled from whatever happens to exist. The spec names the modality, the schema and label definitions, the volume, the acceptance criteria that separate a correct item from a rejected one, and the licence terms. It is agreed before a single item is sourced, and the pilot, the production run, and any standing order are all measured against it.
The reason to commission one is distribution. Public corpora and bulk marketplaces are heavy on the cases every model already handles and thin on the ones yours fails. A custom set is built from your failure cases outward: the domains, formats, edge conditions, and label semantics the next training run has to cover. More data is cheap. The right data has to be made.
eeko's data practice exists because the production systems we build, from enterprise AI agents to retrieval pipelines, run on models that need data nobody has published. The custom datasets we assemble for our own deployments are the same ones offered here: consented, human-made, expert-checked, with a verifiable record on every item.
Custom AI training datasets, spec to receipt.
Six pieces of work. Some engagements need all of them, most need four. Each is scoped against the same specification and priced before it starts.
- Label definitions with worked examples
- Acceptance criteria you can run
- Exclusivity settled up front
- Opt-in creators on stated terms
- Fingerprinted against duplicates
- Origin class on every item
- Credentials verified by field
- Attested human authorship
- Second-expert review on labels
- Representative of the full spec
- Approved item by item
- Priced separately, so you can stop
- Same criteria as the pilot
- Failed items replaced, never delivered
- Signed receipt naming every item
- Scheduled deliveries
- Spec versioned as needs change
- Same receipt, every drop
Two channels, one intake.
Most custom datasets mix existing work for breadth with commissioned work for the hard cases. Whichever channel an item arrives through, it faces the same checks.
How a set gets filled
- Catalogue — photographers, writers, voice talent, and companies who have listed human-made work with AI-training permission and their own terms. Volume, coverage, speed.
- Commission — clinicians, lawyers, engineers, mathematicians, translators, and tradespeople in the eeko expert network, paid to make new items to the spec. The cases nobody has recorded.
- Mix — the usual answer. Catalogue material for the body of the distribution, commissioned material for the tail, both held to the same criteria.
- Intake — fingerprint, origin check, safety and personal-data screen, consent and licence binding, expert review, audit trail. Identical for both channels.
We do not pad a delivery with near-duplicates to hit a number.
What happens to every item before it ships.
One record from a commissioned set of de-identified clinical notes, followed through intake. The steps are the same for a photograph, a recording, or a CAD file.
From failure cases to a signed delivery.
Fixed scope, fixed price, pilot before volume. The same way we run every engagement at eeko.
Custom datasets by field.
What counts as a correct label, which releases are needed, and who is qualified to review all change with the field.
Custom AI dataset questions.
How much does a custom AI dataset cost?
Price follows the specification. The levers are modality, volume, how much is newly commissioned versus drawn from the catalogue, how senior the experts are, how deep annotation and review go, and whether the licence is exclusive. A commissioned clinical set with second-physician review costs more per item than a catalogue-sourced image set. We quote a fixed price against the written spec, with volume pricing on larger runs. The pilot batch is priced on its own, so you can stop after it.
How long does it take to build a custom dataset?
It depends on how much has to be made new and how rare the expertise is. Sets drawn mostly from the catalogue move in days; sets needing commissioned recordings or specialist authors take longer. Dates go into the specification, the pilot ships first, and we send progress against the spec while production runs, so you can plan a training run around a real date. A data lead answers every request within one business day.
What formats are custom AI datasets delivered in?
JSON Lines, Parquet, CSV, WebDataset, COCO, and Croissant metadata. Every record travels with its labels, metadata, origin class, licence pointer, and hash, so nothing needs a second lookup. If your loader expects a particular schema, we write it into the spec and validate each delivery against it before signing. Deliveries arrive as signed bundles with a receipt that names every item by ID and hash.
Can we license a custom dataset exclusively?
Yes. Commissioned work can be licensed exclusively as standard, and catalogue items can be when their creator has opted into exclusive terms. The exclusivity flag lives on the item itself, so a work licensed to you for a given scope is not offered to anyone else for that scope. Because it moves both price and sourcing time, we settle exclusivity in the specification, before the pilot rather than after.
How do you check the quality of a custom dataset?
Against acceptance criteria written into the spec and proven on the pilot batch, which you approve one item at a time. Production is held to the same bar. Each label is tagged with its source, whether contributor, model suggestion, or expert reviewer, and the spec fixes what share of items a second expert re-checks. Anything that misses the criteria is replaced rather than delivered.
What happens to items that fail verification?
They never reach you. An item that duplicates or conflicts with registered work, lacks consent or a required release, trips the safety or personal-data screen, or carries generator signatures its contributor did not declare is held back and replaced with one that passes. The delivery receipt lists only items that cleared every check, so what you trained on and what was verified are the same list.
Is the data in a custom dataset human-made or synthetic?
You decide, and the record proves it. Every item carries an origin class: human-made, AI-assisted, or AI-generated. Contributors attest to how the work was produced, and intake scans metadata for generator signatures and C2PA or IPTC assertions to test the claim. If the spec says fully human-made, assisted and generated items are excluded. If you want a controlled share of synthetic material, the classification lets you set the mix and audit it later.
Related data services.
Bring us the cases your model gets wrong.
Five minutes to send a dataset request. Within one business day a data lead comes back with questions, a draft specification, and a price. No deck. A ranked list of what to source first.
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: