AI Training Data Glossary: Licensing & Provenance Terms

AI data lab: An organization that sources, creates, labels or evaluates data used to train and test AI models. Some operate expert networks, some run annotation workforces, and some license existing datasets from third parties. Read more
Chain of title: The record showing that each party in a chain had the right to pass rights to the next. Gaps in chain of title are one of the most common reasons a buyer declines a dataset.
Data broker / intermediary: A party that connects data owners with buyers, sometimes taking a fee or reselling licenses. Not every intermediary takes title to the data.
Data room: A controlled environment where a prospective buyer can inspect samples or documentation under NDA before a license is signed.
Datasheet / dataset card: A structured document describing a dataset's purpose, composition, collection process, licensing and known limitations. Read more
De-identification: Removing or transforming identifiers so records cannot reasonably be linked to individuals. Legal standards differ by jurisdiction; it is not the same as anonymization in every regime.
Evaluation data: Held-out data used to measure model performance. It is often smaller than training data but must be clean, well-documented and not leaked into training sets.
Exclusive license: Permission that the owner agrees not to grant to anyone else, usually within a scope (field, territory, time). Generally commands a premium and narrows future options.
Field of use: Contractual limit on what a licensee may do with data, e.g. 'training and evaluation of speech models' or 'internal research only'.
First-party data: Data an organization collected or created itself, with direct knowledge of how it was produced and what consents apply.
Non-exclusive license: Permission to use data that the owner can also grant to others. Typically lower price per buyer but repeatable.
PII / personal data: Information relating to an identifiable person. Datasets containing personal data need a lawful basis and often consent or de-identification before licensing.
Provenance: The documented origin and chain of custody of data: who created it, how it was collected, and every transfer of rights since. Read more
RLHF: Reinforcement learning from human feedback: a training technique that uses human preference judgments, often from domain experts, to steer model behavior.
Sample set: A small, representative, rights-cleared extract shared for evaluation. Should never include material you are not authorized to share.
SFT (supervised fine-tuning): Fine-tuning a model on curated input–output examples. Expert-written demonstrations are a common SFT data type.
Synthetic data: Data generated by models or simulations rather than collected from the world. Its value depends on fidelity and on the rights to whatever seeded it.
Training data: Examples a model learns from during pre-training or fine-tuning. Buyers care about quality, relevance, diversity and whether they have the right to use it. Read more

Continue reading