AI data glossary

This glossary defines 35 terms used in AI training data and data auctions. Each entry opens with a one-sentence definition and links to related terms and deeper guides.

Last reviewed November 7, 2026

AI data auction: An AI data auction is a competitive sale in which buyers bid, or sellers undercut, to set the price and terms for training or evaluation data used by AI models.
annotation: Annotation is the process of adding labels, ratings or structure to raw data so a model can learn from it.
data broker: A data broker is a business that collects data from many sources and sells or licenses it to others.
data card: A data card is a short document that describes a dataset's contents, collection method, limits and intended use.
data clean room: A data clean room is a controlled environment in which parties analyze or train on combined data without exposing the raw records.
data exchange: A data exchange is a marketplace built into a cloud or data platform so customers can access third-party data within their own accounts.
data licensing: Data licensing is the grant of rights to use a dataset for stated purposes, in return for payment or other terms.
data marketplace: A data marketplace is an online platform where data owners list datasets and buyers search, compare and license them.
data provenance: Data provenance is the documented history of where data came from, who created it and what rights attach to it.
evaluation set: An evaluation set is a collection of tasks used to measure a model, kept separate from training data.
expert demonstration: An expert demonstration is a recording of a skilled professional performing a task step by step, used to teach a model the process.
fair use: Fair use is a US copyright doctrine that permits some uses of copyrighted work without permission.
fine-tuning: Fine-tuning is the further training of a pretrained model on a smaller dataset to adapt it to a task or style.
foundation model: A foundation model is a large model trained on broad data that can be adapted to many tasks.
frontier lab: A frontier lab is an organization that trains the most capable general-purpose AI models.
held-out set: A held-out set is a portion of data deliberately excluded from training so it can test the model fairly.
inter-annotator agreement: Inter-annotator agreement is a measure of how often different reviewers give the same label to the same item.
neolab: A neolab is a newly founded, well-funded AI research company that is not one of the established frontier labs.
opt-out: An opt-out is a signal or process by which a rights holder tells AI developers not to use their content for training.
PII: PII, or personally identifiable information, is any data that can identify a specific person.
post-training: Post-training is the set of steps after pretraining, such as supervised fine-tuning and reinforcement learning, that shape a model's behavior.
preference data: Preference data is a collection of human choices between model outputs that records which response people prefer.
pretraining: Pretraining is the first and largest stage of training a model, in which it learns general patterns from a very large body of text, code or media.
reserve price: A reserve price is the minimum price at which a seller will complete an auction.
revenue share: Revenue share is a payment model in which a data owner receives a percentage of the revenue generated from using the data.
reviewer calibration: Reviewer calibration is the process of aligning reviewers' judgments with a gold standard through examples, feedback and checks.
reward hacking: Reward hacking is a failure in which a model finds a shortcut that earns reward without doing the intended task.
RL environment: An RL environment is a simulated task with tools and an automatic grader in which a model practices and earns rewards.
RLHF: RLHF, or reinforcement learning from human feedback, is a training method in which human preference judgments are used to train a reward model that then guides a language model.
rubric: A rubric is a written set of criteria that tells reviewers how to score or compare a model output.
SFT: SFT, or supervised fine-tuning, is a post-training method in which a pretrained model learns from prompt and ideal-response pairs.
synthetic data: Synthetic data is data generated by a model or simulator instead of collected from real people or events.
training data: Training data is the set of examples a machine learning model learns from.
Vickrey auction: A Vickrey auction is a sealed-bid auction in which the highest bidder wins and pays the second-highest bid.
winner's curse: The winner's curse is the tendency of the winning bidder in an auction to overpay because they had the highest, and often most optimistic, estimate.