Terms that describe the collection
A dataset is a defined collection, not necessarily an entire source system. A record is one structured entry; an example is a unit chosen for a task. A conversation or video clip may contain many records but count as one example. A corpus usually refers to a collection of text or related content.
A schema describes structure and fields. A data dictionary explains their meanings. A manifest lists the items in a delivered version. Provenance describes origin and history. These documents answer different questions and should not be treated as interchangeable attachments.
Terms that describe model use
Training changes a model using data. Evaluation measures behavior on a task. Fine-tuning is additional training of an existing model. Retrieval supplies relevant material to a system at use time; it is not automatically the same as changing model weights. Ask the buyer which activities are actually proposed.
Annotation adds information such as labels, transcripts, or ratings. Synthetic data is generated rather than directly captured from the original event, but it still needs a documented process and rights review. A benchmark is an evaluation setup, not merely any collection with answers attached.
Terms that describe permissions
A license grants specified permissions; it does not necessarily transfer ownership. Exclusive and non-exclusive describe whether competing grants are restricted within a defined scope. Sublicensing allows a licensee to grant some rights onward, subject to the actual contract. A field-of-use restriction limits the application rather than necessarily the recipient.
De-identification is a broad term whose meaning depends on context. Do not use it as an automatic promise of anonymity. Pseudonymisation replaces or separates identifying information but can leave data within privacy requirements. The ICO’s guidance explains why this distinction matters in its legal framework.
- Specify evaluation and training rights separately.
- Define the accepted unit for any per-unit price.
- Ask what rights continue after raw-data access ends.
Terms that describe delivery and payment
Acceptance criteria are the agreed checks used to decide whether a deliverable meets requirements. A milestone ties a defined event to a responsibility or payment. A refresh is a later delivery of updated or additional material. Retention describes how long data or artifacts are kept, which may differ from the commercial license term.
When a term is ambiguous, replace it with a concrete example in the negotiation. Ask what the buyer can do with one approved record, who can see it, and what happens after expiry. Plain definitions improve discussions, but the signed agreement controls the actual bargain and should receive appropriate review.
Common questions
Are tokens a reliable way to compare every dataset?
No. They are useful for some text-processing measurements, but do not capture rights, uniqueness, audio duration, visual coverage, or task completeness.
Does proprietary mean cleared for sale?
No. The word is often used loosely. Ask what ownership, licensing, privacy, and confidentiality evidence supports the proposed transaction.