AI training data glossary

The first, largest training stage, in which a model learns from very large amounts of general text, code or media. Most disclosed content licences (news archives, forums, books) feed this stage.

See: Deals database

Further training on worked examples: an input and the answer a capable person would give. A question in a support queue and the reply that resolved it has the same shape.

Training a model by having people rank or grade its outputs, then rewarding the outputs they prefer. The people doing the grading are often paid experts.

Example: Mercor lists software engineering experts at $70 to $150 an hour for grading model-written code. (Mercor job listing, checked 18 Oct 2026)

A simulated workplace in which an AI agent attempts real tasks and is scored on the result. Built from records of how work was actually done, so the agent can practise the same steps.

Example: Surge AI uses startup archives to build reinforcement-learning gyms for training AI agents. (MIT Technology Review, checked 18 Oct 2026)

An anonymised replica of a company's tools and records, used as the setting for an RL environment.

Example: Ooak Data describes turning Gmail, Slack, Notion and Jira data into an anonymised digital twin for agent training. (Y Combinator, checked 18 Oct 2026)

Examples held back from training and used only to test a model. Because they must stay unseen, eval data is often licensed under stricter limits than training data.

Training material written or graded by specialists such as engineers, lawyers or doctors, usually paid per hour rather than per item.

See: Expert hourly rates in the price table

A licence gives the buyer the right to use a copy on agreed terms while the owner keeps the original. A sale transfers the asset itself. Most programmes describe licences; SimpleClosure offers either.

See: SimpleClosure profile

A licence that bars the owner from licensing the same material to anyone else, for a period or for good. It usually commands a higher price.

Example: Pewter lists some deals as a perpetual, exclusive licence. (Pewter, checked 18 Oct 2026)

A licence that leaves the owner free to license the same material again. Common in large content deals.

Example: Calliope Networks says its AI model training licences are non-exclusive. (Calliope Networks, checked 18 Oct 2026)

A licence with no end date. The buyer may use the material, and keep what a model learned from it, indefinitely.

Payment as a percentage of what an intermediary earns each time it licenses the data again, instead of or on top of an upfront sum.

Example: Replay pays upfront cash plus a perpetual revenue share; Troveo pays on every resale. (Replay, checked 18 Oct 2026)

Payment due 30 or 60 days after a trigger such as invoice or approval.

Example: Polyshares pays Net 30 to 60 once data is approved, shared and anonymised. (Polyshares, checked 18 Oct 2026)

A clause that lets the payer recover money already paid, for example if the data turns out incomplete or a downstream buyer rejects it.

See: Licence clauses compared

A figure shown before anyone has examined the records. It binds no one and is not an offer.

Example: Handshake AI labels its on-page estimate indicative and not an offer to purchase. (Handshake AI, checked 18 Oct 2026)

Money a programme pays to whoever introduces a company that goes on to license data.

Example: Grepped pays 10% of the deal, up to $100K per referral. (Grepped, checked 18 Oct 2026)

A marketplace lists data for buyers to purchase; a broker places it with a buyer and is paid a fee; a direct buyer licenses it for itself and resells or uses it.

See: Programme directory

Anything that identifies a person directly, such as a name, email address or phone number, or indirectly in combination with other details.

Replacing identifiers with placeholders while a key to reverse it exists somewhere. Under the GDPR, pseudonymised data is still personal data.

Processing that leaves no reasonable way to identify anyone, with no key kept. Truly anonymous data falls outside the GDPR, which is why the distinction matters in licensing.

The umbrella term for removing or replacing identifiers before records leave a company. Programmes use it loosely; ask which of the two above they mean.

How long a software vendor keeps a customer's data after cancellation or non-payment before deleting it. Exports often stop working before deletion starts.

See: SaaS retention reference

A price per item: per word, per image, per minute of video, per book. Common for media, rare for company records, which are priced per company.

See: Unit prices