Skip to main content

Menu

How to buy AI training data

You can license an existing dataset, work directly with a data owner or commission a collection for a specific task. Before committing, define what the model needs to learn, test a sample against that requirement and check whether the proposed terms cover the intended use.

The useful question is not only “How much data is available?” It is whether the records contain the examples and context your team can use, and whether the agreed license covers the people, systems and purposes involved.

Choose a sourcing route

RouteWhen to consider itWhat to ask first
Existing datasetYou need a collection that may already exist.Which version, fields, time period and usage terms are included?
Direct agreement with a data ownerOne organization holds relevant records or content.Who can authorize the proposed use, and what material is in scope?
Custom collection or preparationExisting material misses important scenarios, formats or labels.What will be collected or prepared, how will quality be checked, and what can you use afterward?
Public datasetAn accessible collection may fit the task or provide an initial comparison.What does its documentation say about source, limitations and permitted use?

For example, Opendatabay presents a marketplace of dataset listings, while Clickworker describes custom collection and annotation services. These are starting points to investigate; check the offering and terms for the particular data you need.

Our data sourcing guide covers how to describe the task, coverage and format before approaching a source.

Define what a useful example contains

Write down the intended behavior and the information needed to assess it. For a support agent, a record might need the customer's request, the policy that applied at the time, the action taken and the outcome. For a document model, the important fields might be page images, extracted text, layout and verified labels.

If a document sample includes labeled boxes, check that the boxes align with the delivered images. Some formats express positions as fractions of page dimensions; others use pixels. Confirm the page, image version, units, coordinate origin and any resizing, cropping or rotation, then overlay sample boxes to check alignment. Also review whether the labels describe the right regions.

Ask for a description of the exact dataset version: its source, size and units, time coverage, formats, labels, known exclusions and processing history. Hugging Face's dataset-card guidance describes how a card can document contents, context and metadata such as license, language and size. Treat those fields as information to review alongside the underlying documentation.

Test a sample against your requirements

Agree on a sample and the checks before deciding whether to expand. Include the routine situations your model will encounter and the difficult cases that matter to the task. Record how the sample was selected and which parts of the proposed collection it represents.

Origin Lab's Marketplace Terms, read on 18 November 2026, restrict its defined marketplace samples to internal evaluation and prohibit using them to test or evaluate an AI system unless an applicable commercial agreement executed by Origin Lab expressly states otherwise. That sample restriction excludes separately defined Bench Materials, which have their own terms. Check the sample's specific terms and any agreement before inspecting, testing, training or sharing it.

For a speech sample with word timestamps, ask which recording and time unit they use. If the timestamps refer to a copy with silence removed, ask for the source recording identifier and a map of the retained spans to the original timeline. Check a word after an interior cut. In a fictional sample that joins source seconds 2–4 and 7–9 without padding, 2.5 seconds in the sample maps to 7.5 seconds in the source. Adding the first span’s two-second offset gives 4.5 seconds, inside the removed gap. NeMo Curator timestamp mapping

Check how the delivery marks words whose timestamps cross a join. In the same fictional sample, the tested faster-whisper helper chose one offset for both ends of a crossing interval and placed its start within the removed gap. If your use requires a continuous interval in the original recording, agree how to flag and review those cases before accepting them. faster-whisper timestamp restoration · faster-whisper offset calculation

Consider this fictional example. A team wants to fine-tune a support agent that recommends the next step for a return request. A supplier proposes a collection of support cases. The team asks for 12 diagnostic examples: six routine returns and six exceptions. For this first review, it requires every example to include the customer request, a link to the policy version that applied and the recorded outcome.

CheckRequirement for this sampleIllustrative result
Customer requestPresent in all 12 examples12 of 12 have it
Applicable policy versionLinked in all 12 examples10 of 12 have it; two exception cases lack the link
Recorded outcomePresent in all 12 examples11 of 12 have it; one routine case lacks an outcome

The sample fails two of the team's stated checks. The next step is to ask whether the missing policy links and outcome can be recovered and to review a corrected sample. If they cannot, the team can change its requirement or look for another source. A recorded outcome is not automatically a correct decision: a domain reviewer should also compare outcomes with the applicable policy.

These counts describe only the 12 selected cases. They do not estimate the error rate in the full collection. To assess broader coverage, ask how the dataset is distributed across relevant scenarios and agree on a further sample drawn using a documented method. Google's guidance on data quality and sampling explains why a selected sample can give a misleading picture of a larger population.

Suppose those 12 cases all came from 2024, but the offered collection covers 2022–2024. Checking a corrected copy can confirm the fixes to those cases; it does not show whether the rest of the collection was corrected. Ask which dataset version and records received the correction, then agree how to select a fresh sample from the version you intend to buy. If only 2024 can be reviewed, leave 2022–2023 unassessed and decide whether to narrow the purchase or wait for a broader review.

If real records cannot be shared for the initial review, ask whether a sanitized sample or controlled access is possible. A synthetic example may show the format, but it cannot establish the quality or coverage of the actual collection.

Agree on how access is enforced before receiving a sample: a query that selects certain columns does not, by itself, restrict access to the rest of the underlying file. A Parquet footer can also contain column statistics. If the sender needs to limit what crosses the boundary, consider a separately prepared sample with approved fields or a service that enforces field-level access. Review metadata and access controls alongside the returned records.

Review the proposed use and evidence of rights

Match the agreement and supporting documents to the specific dataset and project. Questions to discuss with the supplier and the people responsible for legal and procurement review include:

  • Which dataset version is covered, and is the supplier authorized to grant the proposed use?
  • Are training, fine-tuning, evaluation or retrieval included where needed? Do the terms address commercial deployment and any relevant restrictions on models or outputs?
  • Who may access or process the data, including contractors or service providers? What are the delivery, storage, retention and deletion terms?
  • Are there exclusions, upstream rights, third-party material or other restrictions that need review? What evidence is available for the supplier's account of the data's origin and permissions?

Make a separate note of what is confirmed and what still needs review. A license label or dataset card helps identify a claimed use; it does not by itself verify the underlying rights. An audit of public text, speech and video datasets found differences between dataset-level license descriptions and restrictions associated with source content. Its findings concern the public datasets examined, not every commercial collection.

For example, a pinned card for the public research dataset SciFact describes the dataset as CC BY-NC 2.0 and links to a license file on the repository's master branch, which can change. In a pinned snapshot of that license, the repository separately describes CC BY 4.0 for claims and evidence annotations, ODC-By 1.0 for corpus abstracts, and Apache 2.0 for its code. The documents may describe different versions, components or transformations; they do not by themselves establish the terms for a particular file. Map delivered files and their upstream components to the relevant documentation, and have unresolved rights questions reviewed.

Plan for evaluation and delivery

Decide how the data will be delivered and accepted: file formats, a manifest or version identifier, documentation, sample checks and how problems will be corrected. If the data will also be used to evaluate a model, protect a separate test set from training and check for duplicates across the split. Google's dataset-splitting guidance explains how duplicates can make a test look stronger than it is.

Check when the inspection window starts, how to give notice, what silence means and what remedies apply. For example, version 1.0 of ai.market's standard AI-training licence gives a buyer seven calendar days after first delivery to notify the marketplace of a material mismatch with the listing. micro1's master terms, updated 12 November 2026, give micro1, the licensee, 30 business days after delivery is completed under the manifest to reject data in writing. A separate trigger applies if the supplier notifies it in writing that the remaining data cannot be delivered. The documents apply to different transactions and contain further conditions and exceptions. Review the terms and version that govern your order.

A manifest can list the expected files and checksums for the agreed dataset version. Agree how you will verify the manifest's source, compare its file list with what arrived, and investigate missing, unexpected or mismatching files. For example, if a fictional manifest lists two files but only one arrives, a matching checksum for that file does not make the delivery complete. The BagIt specification distinguishes completeness from checksum validation for one packaging format. A checksum match says that a file's computed digest agrees with the manifest; it does not by itself establish the sender's identity, correct labels, representative coverage or rights to use the data. Review those separately.

For a worked example of checking test-set leakage and missing context, see our dataset evaluation guide.

Sapien1 helps teams scope data requirements, identify possible sources and assess samples. Start with the task, the examples you need and the decision a sample must support. Data owners preparing for a conversation can use our data licensing preparation guide.