How to write a training data request sellers can answer

Most data requests fail before any seller reads the second paragraph. A request that says 'we need high-quality legal data' gets either silence or a sales call. A request that names the record types, the date range and the rights needed gets a quote.

This guide lists the fields a request should cover, what to write in each and why sellers need it. It is the requirement stage of a training data procurement.

What makes a training data request answerable?

A training data request is answerable when a seller can check each field against its own records and reply yes, no or partly. Every field should be a range, a list or a rule, never an adjective.

Start with the training use. Pre-training, supervised fine-tuning, preference tuning, evaluation and reinforcement learning need different data shapes and different rights, so the same archive can carry very different prices depending on the use.

Which fields should a training data request include?

A training data request should include ten fields after the training use. The table shows what to write in each and why the seller needs it.

How do you specify domain and volume?

Specify domain as record types from named source systems, and volume as a minimum and a target in the unit you will price on. Sellers search their archives by system and record type, not by topic.

Pick the unit carefully. Matter files, support tickets, hours of audio, repositories and comparisons each count work differently. If you will pay per unit, the unit in the request becomes the unit in the contract. Our training data catalog lists the usual unit for each data type, such as support tickets or expert domain records.

A smaller, cleaner set can beat a large one. The LIMA paper from 2023 fine-tuned a 65 billion parameter model on only 1,000 curated prompts and responses, and human raters judged its answers equivalent or preferred to GPT-4 in 43% of cases. State a minimum you can use rather than the largest number you can imagine.

How do you specify time range and formats?

Give the earliest and latest creation dates you will accept, and the file types and structure you need on delivery. Dates matter for relevance, for privacy notices in force at collection and for whether records predate widespread AI-generated content.

For formats, say whether you want native files, a converted format such as JSONL, or both. Ask for documentation as a datasheet, a format proposed by Timnit Gebru and colleagues that records a dataset's motivation, composition, collection process and recommended uses, or as machine-readable metadata in the MLCommons Croissant format.

What rights should a training data request ask for?

Ask for the narrowest rights that cover your real use: which uses, which models, which affiliates, what term and what territory. Broad rights cost more and take longer to clear, and a seller's lawyer will ask why you need them.

• Uses: training, fine-tuning, evaluation, or all three
• Models: named internal models, future versions, or models offered to customers
• Derived assets: ownership of annotations, embeddings, evaluation sets and fine-tuned weights
• Onward transfer: whether cloud vendors, subprocessors or model partners may touch the data
• Term and expiry: what happens to trained models and retained copies at the end

Do not treat US fair use as a substitute for the rights you need. US courts have reached different results on training so far, and a license gives you certainty a defense cannot. Our data rights page covers the background.

Ask for exclusivity only when your advantage depends on competitors not training on the same data, as with a private evaluation set. Otherwise ask for a non-exclusive license or a limited exclusive by field, term or territory.

Our guide to exclusive vs non-exclusive data licenses sets out the options and how they change price.

How do you state de-identification requirements?

State the legal standard the data must meet, the identifiers to remove and who will do the work. Sellers price de-identification as real effort, and they cannot quote it against 'anonymized'.

• Health data: HIPAA allows de-identification by expert determination or by the safe harbor method of removing 18 types of identifier, according to the US Department of Health and Human Services.
• California consumer data: Nixon Peabody notes that under the CCPA, data counts as de-identified only if it cannot reasonably be linked to an individual, the holder publicly commits not to re-identify it and recipients are contractually bound.
• Business records: list the identifiers to remove, such as client, counterparty and employee names, and ask for a documented method and a re-identification test on a sample.

Our guide to de-identification for AI training data explains the methods sellers use.

How should you describe delivery?

Describe the delivery method, the environment the data will live in and the schedule. Nixon Peabody notes that for especially sensitive archives, source data can stay in a secure environment where defined training or evaluation takes place, without transferring unrestricted copies.

If you can accept that model, say so in the request. It widens the pool of sellers willing to talk. Also ask for a manifest with record counts and file hashes, and for delivery in tranches if you plan to pilot first.

How do you state a budget model?

State the pricing basis as well as the range: flat fee, per unit, recurring access or revenue share. Sellers can judge viability from a model faster than from a number.

Public filings show how large sellers structure this. Reddit's 2024 S-1 states that the transaction price in its data licensing arrangements is generally a fixed fee or usage-based fee, recognized over the license period, with payment usually due within 30 days. Our guide to AI training data pricing models compares the structures, and the AI data price index lists published ranges.

What should you leave out of a training data request?

Leave out model architecture, roadmap detail and anything you would not want a competitor to read. A first request can go to many sellers, so keep it to the fields above and send detail only under a mutual NDA.

If you hold or buy EU or UK personal data

Add the lawful basis you expect the seller to rely on and whether data will leave the EU or UK. Under the EU AI Act, providers of general-purpose AI models also publish a training content summary, so ask sellers for the source descriptions you will need for it.

How do you send a request through Sell My Data To AI?

Our data requirement form is built around the fields in this guide. Buyers pay nothing to submit a requirement or review a shortlist, and any fee is agreed in writing before an introduction. Every seller we shortlist has passed our eight-point vetting standard.