Skip to content
Sell My Data To AI
Menu

Data sample request checklist for AI training data

A checklist both sides can work from when a buyer asks to see a sample: what to agree first, how to draw it, what to send with it and what happens after.

In short

This data sample request checklist covers what a buyer should ask for and a seller should prepare when an AI training data sample changes hands. It runs from the mutual NDA and named access, through sample size and a stratified random draw, to the documents that go with the sample, the tests the buyer will run, written results and certified deletion. Both sides can work from the same list.

Key takeaways

  • A sample is only useful if it is drawn at random across the whole archive, with the method and random seed recorded, not hand-picked.
  • The sample should go through the same exclusions and de-identification as the full delivery, in the same formats.
  • Sample results should become the written acceptance criteria for the full delivery.
  • Representative sample is one of the eight checks in the Sell My Data To AI seller vetting standard.

A starting point for discussion, not legal advice

This template is written in plain English for US use and makes no claim about how a court would treat it. We are a broker, not a law firm. We tell you what buyers will check and help you prepare, but you should take independent legal advice before signing any license.

The files are built in your browser. Every detail to fill in is shown in [brackets].

How to use this template

  1. 1

    Agree the items in the first section before any data is drawn.

  2. 2

    Fill in each [bracketed] placeholder, and tick items as they are completed.

  3. 3

    Keep the completed checklist with the data manifest and the NDA.

  4. 4

    Turn the test results into acceptance criteria for the license.

The template

Data sample request checklist

Dataset: [Dataset name and manifest version]

Seller: [Seller legal name]

Prospective buyer: [Buyer legal name]

Prepared by: [Name, title, date]

Before the sample is drawn

  • Mutual NDA signed, covering the sample, with no-training and no re-identification terms and a deletion deadline
  • Buyer's identity, signatory and intended use confirmed in writing
  • Data manifest received by the buyer, version [Version] dated [Date]
  • Names of the people who will access the sample: [Names]
  • Storage location and access controls agreed: [Location]
  • Evaluation period and deletion date agreed: [Date]

Sample specification

  • Sample size agreed: [Number of records or share of the archive]
  • Stratified random draw across years, systems and teams, not hand-picked
  • Random seed or selection script recorded so the draw can be repeated
  • Same formats, schema and structure as the full delivery
  • Same exclusions and de-identification applied as the full delivery
  • Linked records kept together, such as threads with replies, tickets with resolutions and code with its commit history
  • File list with a SHA-256 hash for each sample file

Documents sent with the sample

  • Sampling method statement, signed off by [Name]
  • De-identification report: method, tools, review sample size, error rate and re-identification test results
  • Exclusions list, by category
  • Rights summary for each source: who created the records, IP assignments, and the contracts and platform terms reviewed
  • Privacy notice in force when the data was collected, or confirmation that the sample holds no personal information
  • Known issues: duplicates, gaps in coverage and corrupt files

Tests the buyer will run

  • Record counts and date ranges against the manifest
  • Format, schema and parsing checks
  • Scan for personal information and secrets, such as names, email addresses, phone numbers, account numbers and API keys
  • Duplicate and near-duplicate rate
  • Language mix against the manifest
  • Task fitness tests: [Describe each test]

After the review

  • Written results shared with the seller by [Date]
  • Acceptance criteria for the full delivery agreed, based on the sample results
  • Sample deleted and deletion confirmed in writing by an officer of the buyer by [Date]
  • Next step agreed: [term sheet], [paid pilot] or [no further action]

Frequently asked questions

How big should an AI training data sample be?
There is no fixed size. The sample should be large enough to cover every system, year and team in the archive and to run the buyer's tests with confidence, and small enough that losing it would not hurt the seller. The checklist leaves the size as a placeholder for the two sides to agree. Our guide to evaluating dataset samples explains what buyers test.
Why does the sampling method matter?
A hand-picked sample shows the best of the data, not what the buyer will receive. A stratified random sample, drawn across years, systems and teams with the random seed recorded, lets the buyer judge the whole package and repeat the draw. That is why a representative sample is one of the eight checks in our seller vetting standard.
Should a sample be shared before an NDA is signed?
No sample should be shared before a mutual NDA is signed that bars training on it and requires deletion. Our mutual NDA template covers those terms. We do not take custody of your raw data. Buyers see redacted samples under a mutual NDA, and full data moves only after you sign a license directly with the buyer.

Related guides

For data owners

Want a second pair of eyes?

A free valuation gives you a view of likely buyers, a realistic range and the rights issues to clear before a sale. Sellers pay nothing to be assessed.

Get a free valuationBook a free consultation

For AI teams

Tell us what training data you need

Submit a written requirement and we find, vet and shortlist sellers of rights-cleared training data. Buyers pay nothing to submit a requirement or review a shortlist.

Submit a data requirementBook a free consultation