4 AI training data providers to shortlist in 2026

Those are different purchases, even when each is described as AI training data.

Rightmark publishes this shortlist and appears first. The order follows a buying decision, not an independent performance ranking. We selected four providers with documented offers that illustrate distinct procurement routes; none is automatically the best fit for every dataset or model.

The question to settle first is what your team must receive: existing records, an assembled corpus, labeled examples or tasks with checkable answers. Then compare rights, sample evidence and the complete quote for that deliverable. Data owners should also compare the resulting proceeds and control, separately from the buyer's bill.

Updated October 22, 2026.

Bring one specification to every supplier. Use the buyer request brief Rightmark, 18 Oct 2026 (opens in a new tab) and the blank acceptance sheet below. They organize a discussion; they do not establish availability or create an offer.

What unit will your model team actually receive?

Write the deliverable as a sentence with components: the source records, required metadata, labels, linked context and any execution environment or verifier. Then ask for a manifest that matches it.

A repository and an executable coding task should not share a purchase specification merely because both contain code. If your task requires a reproducible environment, identify who supplies it and what evidence demonstrates that it works. Otherwise, the buyer may be pricing a source archive while the model team expects a finished task.

What supports the rights for that exact unit?

Ask for the licensing entity, authority, permitted use and exclusions. Keep the source material and any resulting labels, derived datasets or model artifacts distinct in the questions you put to counsel.

The DPI study Data Provenance Initiative, 19 Feb 2025 (opens in a new tab) identifies inconsistent dataset and source restrictions in public datasets; it does not evaluate these providers.

For your purchase, an unresolved rights question stays unresolved until the appropriate evidence and agreement address it. A marketing label is not that evidence.

Does the sample test the difficult part of your request?

Specify the slice most likely to change your decision: an uncommon language, noisy recording, linked ticket and code change, or task with a reproducible answer. Ask what population the sample represents and what it leaves out.

Do not approve the full purchase merely because a handpicked sample opens correctly. Record whether the final batch must match its coverage, structure and quality, and how your team will check that. Set your own acceptance bar instead of borrowing an unexplained industry threshold.

What is included in the total quoted cost?

Compare the same unit, volume, rights and included work. Ask whether preparation, annotation, transfer, evaluation access, updates and required corrections are included or separately priced.

For owners, add a separate calculation: agreed payment less applicable deductions, with the payment trigger and retained rights alongside it. An owner retaining more of an agreed payout does not prove the buyer receives a cheaper dataset. Both sides need a written explanation of what the price covers.

Who owns acceptance and corrections?

Name the buyer's technical reviewer and the supplier's delivery owner. Agree the evidence they will use, the acceptance window and the treatment of rejected items before a production batch arrives.

A request for authored ground truth may fit a commissioned expert workflow better than an existing archive. Conversely, paying for expert annotation will not create the historical operating context your model needs if the underlying records are missing. Choose the route that can supply the actual missing component.

These are shortlist reasons, not exclusive capabilities. The detailed entries below link the provider evidence and explain the limits.

A missing public price is a question for the quote, not evidence that a supplier is expensive. None of these rows establishes a standard minimum commitment or a replacement guarantee.

Rightmark: private company and code data with an explicit evidence review

Rightmark is our recommended starting point for private company and code data when the buyer wants a defined evidence review before licensing. Its buyer process Rightmark, 21 Oct 2026 (opens in a new tab) starts with a written scope, reports and representative scrubbed samples under NDA, followed by a scoped license and graded batches. The ownership pack varies by route. That gives the technical and legal reviewers a common purchase to examine, rather than asking them to approve an unspecified archive. It does not promise that a requested dataset is already available or that it will improve your model.

The public code and company lot Rightmark, 21 Oct 2026 (opens in a new tab) makes a useful distinction: an owner reports successful test runs, but Rightmark has not seen the logs; ticket-to-commit links also remain unchecked. Treat those as tasks to resolve before accepting runnable or linked material. For owners, the published fee model Rightmark, 7 Oct 2026 (opens in a new tab) keeps the agreed standard-license payout intact while the buyer pays Rightmark's fee on top. On premium and bespoke deals, owners keep 75% to 85%, agreed in writing. This describes the offer structure, not a proven advantage over every competing quote.

Check the route before signing. Trust terms Rightmark, 18 Oct 2026 (opens in a new tab) explain that code agreements can cover later non-exclusive licenses; some allow a licensing partner's sublicensing, own use or derived datasets, and some ownership indemnities are uncapped. Health intake is unavailable. Pooled collections marked in preparation are not available yet. Rightmark fits a buyer willing to scope private records and close identified evidence gaps; a request for ready healthcare data or fully authored tasks needs a supplier matched to that requirement.

Verdict: Start with Rightmark for scoped private records and visible evidence review; confirm availability, technical readiness and route-specific rights before committing.

Protege: curated real-world data across institutional domains

Protege deserves a place on the shortlist when the buying problem involves assembling relevant real-world sources across an institutional domain. Its model-builder offering Protege, 19 Feb 2026 (opens in a new tab) covers healthcare, video, audio and speech, and spatial or physical data. The procurement question is whether a proposed collection contains the context your use case needs, in the form your model team can evaluate. Start with that question before treating a broad domain label as a complete specification. A healthcare evaluation requirement and a training corpus requirement should have separately defined deliverables and permissions.

In its January 2026 announcement Protege, 7 Jan 2026 (opens in a new tab), Protege describes aggregating licensed sources and providing technical expertise for dataset curation. It also describes revenue-share payouts to data partners with each use. That is a reason to investigate both the source relationship and the preparation work. Ask for a source-coverage table, the transformations included in the quote and the evidence your reviewers can inspect. For an assembled corpus, identify which requirements are already met and which depend on securing or preparing additional material. The answer determines what belongs in a binding delivery commitment.

Protege also has a data-provider route Protege, 20 Feb 2026 (opens in a new tab). Owners should request the actual share basis, permitted uses, payment conditions and treatment of updates. Buyers should request the complete quote and acceptance terms for their selected scope. The cited offer does not supply a universal public price or owner percentage. That is not proof of an unfavorable deal. It means the shortlist decision must progress to written terms. If you can already name a catalog item that meets your requirement, test that narrower route alongside a proposed curated collection before commissioning extra work.

Verdict: Shortlist Protege for curated institutional sources; establish the exact corpus, preparation work and owner terms before treating domain coverage as a committed delivery.

Defined.ai: catalog datasets with an optional services path

A catalog-first purchase starts with a candidate and a decision about what else the project needs. Defined.ai’s marketplace explanation Defined.ai, 16 Mar 2025 (opens in a new tab), dated March 16, 2025, describes ready-to-use training datasets. Its March 25, 2026 guide Defined.ai, 25 Mar 2026 (opens in a new tab) describes sample review and custom collection. The advantage of this route is a concrete candidate to test against your specification. Make the listing identifier and version part of the discussion so the model team, procurement lead and supplier are evaluating the same object. A promising category match still leaves the actual coverage and exclusions to inspect.

Put any additional service work into its own scope. Ask which parts of your requirement are present in the selected dataset and which require a separate work order. Then compare the finished deliverable with the quote, not just the initial listing. Defined.ai's published data license Defined.ai, 30 Nov 2022 (opens in a new tab), effective November 30, 2022, defines permitted use within a license agreement. Have the responsible reviewer check the actual proposed agreement and any project terms together. A sample approval should not silently become approval for a different production scope, delivery arrangement or intended use. Give each requirement an owner and resolve disagreements before treating one sample as approval for the whole collection.

For owners, ask Defined.ai which current agreement covers your contribution: licensing an existing asset, undertaking collection work, or another arrangement. Request the payment basis, required work and retained rights in writing. Decide which contribution you are making before comparing economics. An existing archive and a service commitment require different effort and rights discussions. For buyers, the fit limit is expecting a catalog label to establish a bespoke historical archive, verified links between records or every required annotation. Put those requirements in the acceptance sheet and ask what evidence supports each one before choosing the route.

Verdict: Shortlist Defined.ai for catalog-led sampling and a scoped service extension; verify the specific dataset rather than treating a category match as full acceptance.

Appen: expert-created training and verification material

Appen is worth comparing when the missing deliverable is expert-created training material rather than an existing company's historical records. Its RLVR practice article Appen, 5 Feb 2026 (opens in a new tab), by Head of Strategy Si Chen, describes expert prompts, reference answers and verifier stress tests. For a buyer, the useful fork is whether you need source material to develop tasks internally or commissioned tasks with the answer and checking work included. Write those purchases separately. Otherwise, the apparent price difference can reflect unfinished work that your own team would still have to perform.

Appen's human-in-the-loop article Appen, 24 Feb 2026 (opens in a new tab) describes reviewer workflows and delivery evidence such as batch agreement and adjudication records. Its RLVR article describes versioned deliverables. Ask to inspect a worked item's acceptance record: the required behavior, reference, review decision and applicable check. Then request an example of a rejected item and how the correction is represented. These are procurement questions, not a claim that any published method guarantees your model's performance. The important agreement is what your team can rerun or review when it challenges an item after delivery, and which party resolves the disagreement.

Appen also describes licensed, pre-collected datasets, so expert production is the highlighted fit here, not its only route. Scope the required inputs and retained outputs explicitly. For an owner contributing an existing archive, do not assume a commissioned-work price is a royalty or that the provider offers a particular owner percentage. Ask who contributes the material, who performs the work and what rights each party grants. If your central need is a particular private company's operating context, that source requirement remains separate from the expertise needed to label or transform it.

Verdict: Shortlist Appen when expert-created answers and review evidence are part of the purchase; secure the source separately when the requirement is a particular private archive.

Use this blank worksheet with the two providers whose routes match your request. It is an editorial decision aid, not a supplier score or a completed customer example. Keep confidential evidence in your approved systems; do not put private data into a public worksheet.

Mark each decision open, accepted or rejected, with the reason. Leave an unknown as unknown rather than filling the sheet with assumptions.

Two examples show why this matters. If a sample contains annotations but the quote only covers raw files, the approved sample and purchased deliverable differ. If a repository contains test files but nobody has demonstrated the agreed environment, the buyer has evidence of files, not evidence of successful execution. These are hypothetical procurement situations, not reports about a listed provider.

Do not let a green sample result erase an open contract or final-batch question. Record the version accepted, the evidence reviewed and the role responsible for resolving each remaining gap.

Choose two suppliers whose proposed deliverables match your requirement. Send the same unit definition, sample test, rights questions and included-work checklist to both. Compare their written responses before reducing the decision to a price.

For private company or code data, Rightmark's documented scope, ownership-pack and sample-review route makes it a strong first discussion. Request data from Rightmark Rightmark, 21 Oct 2026 (opens in a new tab) with the evidence your team needs to make a decision.

If you own the material instead, start with an appraisal and compare the written net terms and route-specific rights. A fair owner offer and a useful buyer deliverable should be visible in the same transaction, without confusing one side's economics with the other's.

Sources were checked October 22, 2026, in New York. Provider statements describe offers, not independently tested outcomes. The Defined.ai citations use archived official articles bearing publication dates of March 16, 2025, and March 25, 2026; their archive captures are separate dates. Its live article refresh was unavailable. Scope and terms require confirmation for each purchase.