Sell Training Data

What counts as training data

The term covers text corpora, images, audio, video, sensor data, and the annotations attached to any of them. Buyers use it to mean "the material a model learns from", which includes raw content, labeled content, and increasingly the evaluation sets used to test models after training.

Before anything else, check which of these you actually hold. Most suppliers who arrive with "training data to sell" hold one of three things: public data, service capability, or a genuine collection. Only the third is a product.

Why free data is not a product

A large share of the material that could be called training data is already free. Web text is crawled at scale, public corpora are released openly, and donated speech projects cover many languages. If what you hold is a copy of something public, it has no market: buyers who want it can take it themselves, and they know where it is.

This is the part of the market that is genuinely commoditized, and it is where most "how do I sell my data" disappointment comes from. The data was never the asset. What has value is data that is not publicly available and cannot be assembled by anyone with a crawler.

What is actually sellable

Three kinds of supply move in this market. Licensed content with a documented rights chain: material someone owns and can grant use of, where the paperwork is the product as much as the content. Niche collections that match a live specification: rare languages, unusual conditions, specific domains. And annotated data, where the labels were produced by people with the skill the buyer needs.

All three share one property: they cannot be got for free. That is the test to apply to anything you are considering offering. If a competent buyer could assemble it themselves in a week, it is not a product.

The most common reason a supplier is turned away in this market is not quality. It is rights. If you cannot document that you own the material or hold consent for the people in it, it cannot be bought, at any volume and at any price.

The part that is really a service business

A large share of what gets called "selling training data" is actually selling data work: collecting to a specification, annotating, evaluating, quality-checking. This is where most supplier revenue lives, because demand for execution is steadier than demand for any particular file.

Two rules govern the service side. First, data is bought against a specification (a defined volume, format, quality bar, and date), and work that matches no requirement cannot be invoiced. Second, money follows acceptance: the buyer's QA decides whether a batch is paid, and the criteria are set before production, not after.

The exclusions apply here as everywhere: recorded telephone calls, medical or clinical data, and scraped personal data are declined. And no public submission portal exists at any major lab. OpenAI, Google, and Anthropic buy through data teams and vendors, not from uploads, and anyone charging you to submit data to them is selling a story.

What a buyer is actually paying for

When a buyer compares two quotes for the same volume and language, the difference is almost never the audio or the text. It is what comes with it.

The files themselves are the smallest part of the cost. Everything else in the list below is work that has to be done by someone, and a supplier who has already done it can charge for it.

• Recruitment. Finding and scheduling speakers or contributors who genuinely match the required profile, in the required location.
• Consent. Paperwork that names the use, survives a legal review, and can be handed to the buyer's own customer.
• Protocol. A recording or collection method that is consistent across contributors, so the dataset behaves as one thing rather than hundreds.
• Quality control. A defined check, applied to every batch, with a record of what was rejected and why.
• Documentation. Metadata that lets the buyer filter, split and audit the set without going back to the supplier.

What is not sellable

Being clear about this saves a lot of wasted effort. These categories come up regularly and none of them can be sold to a commercial buyer.

The common thread is that the supplier cannot show a right to license the material. Buyers of training data are exposed when the provenance is unclear, so they do not take the risk at any price.

• Anything recorded without the contributor's knowledge or consent. This includes material a contributor has since forgotten agreeing to.
• Recorded telephone calls. We do not take them on and most buyers will not either.
• Clinical or medical recordings, and anything drawn from a patient relationship.
• Data scraped from platforms whose terms prohibit redistribution.
• Datasets assembled from other people's datasets, with no traceable source.