AI training data glossary
32 terms you will meet when licensing or buying data for AI, defined in plain English.
A
- Agent trajectory
- An agent trajectory is the step-by-step record of an AI agent or person completing a task, including each observation, action, tool call and result.
- AI training data
- AI training data is the text, code, images, audio, video or interaction records used to teach a machine learning model, adjust its behavior or test it.
- Anonymization
- Anonymization is a process that removes the association between a dataset and the people it describes, so that no one can be identified from it.
C
- Chain of title
- Chain of title is the unbroken record of ownership that shows how rights in a dataset passed from its creators to the party now licensing it.
D
- Data broker
- A data broker is an intermediary that connects data owners with buyers. In AI training data, a broker values, vets and introduces rather than reselling.
- Data license
- A data license is a contract in which the owner of a dataset permits another party to use it for defined purposes while keeping ownership of it.
- Data manifest
- A data manifest is a structured inventory of a dataset that lists its sources, record counts, date ranges, formats and known exclusions.
- Data marketplace
- A data marketplace is an online platform where sellers list datasets and buyers browse, sample and license them, often with the platform taking a commission.
- Data processing agreement
- A data processing agreement is the contract that limits how a vendor may handle personal data on another company's behalf, such as a CCPA service provider contract.
- Data provenance
- Data provenance is the documented record of where a dataset came from, who created and owned it, and how it was collected, changed and transferred.
- Data room
- A data room is a secure online space where a seller shares documents and samples with vetted buyers during due diligence, with access controls and activity logs.
- Dataset card
- A dataset card is a short document published with a dataset that describes its contents, origin, intended uses, limitations and license.
- De-identification
- De-identification is the process of removing or altering information that identifies people in a dataset, so records cannot readily be linked to individuals.
E
- Evaluation dataset
- An evaluation dataset is a set of tasks with expected answers and scoring rules, used to measure how well a model performs rather than to train it.
- Exclusive license
- An exclusive license gives one buyer sole rights to use a dataset for the licensed purpose, so the seller cannot license the same data to anyone else.
H
- Hold-out set
- A hold-out set is a portion of data kept out of training so it can test how well a model performs on examples it has never seen.
I
- Indemnity
- An indemnity is a contractual promise by one party to cover the other's losses if a named risk occurs, such as a third-party claim over licensed data.
L
- Lawful basis
- A lawful basis is a legal ground under Article 6 of UK GDPR or EU GDPR that an organization needs before it can use personal data.
M
- Mutual NDA
- A mutual NDA is a non-disclosure agreement in which both parties promise to keep each other's confidential information secret and use it only for an agreed purpose.
N
- Non-exclusive license
- A non-exclusive license lets a buyer use a dataset while the seller stays free to license the same data to other buyers on separate terms.
P
- Perpetual license
- A perpetual license grants rights to use a dataset with no end date, so the buyer keeps those rights after any payment period finishes.
- Personal data
- Personal data, called personal information in most US laws, is information that identifies, relates to or could reasonably be linked to a particular person or household.
- Preference data
- Preference data records which of two or more model outputs a person judged better for a given prompt, and is used to train reward models or tune models directly.
- Pseudonymization
- Pseudonymization is processing personal data so it can no longer be linked to a specific person without additional information that is kept separately and secured.
R
- Re-identification risk
- Re-identification risk is the likelihood that people in a de-identified or anonymized dataset could be identified again, alone or by combining it with other information.
- Representative sample
- A representative sample is a subset of a dataset drawn so its mix of record types, dates and quality reflects the whole archive, not hand-picked highlights.
- Revenue share
- A revenue share pays the data owner an agreed percentage of the income a dataset earns, instead of or alongside a fixed upfront fee.
- RL environment
- An RL environment is a software setting in which an AI agent takes actions, receives observations and gets a reward, so it can be trained or tested by trial.
- RLHF
- RLHF (reinforcement learning from human feedback) trains a model on human judgments of its outputs, usually through a reward model learned from preference comparisons.
S
- Success fee
- A success fee is a charge a broker or adviser earns only when a deal completes, usually set as a percentage of the deal value.
- Supervised fine-tuning
- Supervised fine-tuning is further training of a pre-trained model on example inputs paired with target outputs, so it learns to produce the desired responses.
- Synthetic data
- Synthetic data is artificially generated data, produced by models, simulations or rules, that imitates the patterns of real data without being recorded from real events.
For data owners
Ready to see what buyers will offer?
Book a consultation and we will scope your data, explain our fee in writing and set out the steps to competing offers from vetted buyers.
For AI teams
Ready to source?
Book a consultation to brief us. We shortlist vetted sellers, each with a vetting file covering rights evidence, provenance and a representative sample.