AI training data glossary
The terms you'll meet when buying or licensing AI training data, defined in one sentence each.
- Agent trajectory
- A step-by-step record of actions taken to complete a task in software, including screens, inputs and results.
- AI training data
- The examples a machine learning model learns from, such as text, images, audio, video, code, actions and human judgments.
- Anonymization
- Processing data so no one can identify the people in it by any reasonably likely means.
- Benchmark contamination
- When test or benchmark data ends up in a model's training data, inflating its scores without improving real ability.
- Chain of custody
- The record of every party that has held a copy of a dataset, and when, from creation to the current holder.
- Data broker
- In privacy law, a business selling personal data about people it has no relationship with; in AI data, an intermediary matching data holders with buyers.
- Data license
- A contract granting a named party defined rights to use a copy of data, for set purposes and terms, while the licensor keeps ownership.
- Data manifest
- An inventory describing a dataset's contents, sources, volumes, date ranges and rights status, without containing the data itself.
- Data provenance
- The documented history of a dataset: who created it, how it was collected, under what rights, and what was done to it since.
- Dataset card
- A structured document describing a dataset's purpose, composition, collection, processing and intended uses.
- De-identification
- Removing or replacing information that identifies people so data cannot reasonably be linked back to them under a stated standard.
- Evaluation dataset
- A set of tasks with checked answers used to measure a model's ability, kept separate from training data.
- Exclusive license
- A license under which only the licensee may use the data for the agreed scope and term, so the owner cannot license it to others.
- Expert determination
- A HIPAA de-identification method where a qualified expert certifies that the risk of re-identifying individuals is very small.
- Fair use
- A US copyright doctrine allowing limited use of copyrighted material without permission, judged case by case.
- Fine-tuning
- Further training a pre-trained model on a smaller, targeted dataset to improve it at specific tasks or domains.
- HIPAA safe harbor
- A HIPAA de-identification method that removes 18 specified types of identifiers from health information.
- Hold-out set
- Data deliberately kept out of training so it can test whether a model generalizes to examples it has not seen.
- Model collapse
- Degradation in model quality when models are trained repeatedly on data generated by other models instead of human data.
- Non-exclusive license
- A license that lets the owner grant the same rights to other licensees.
- Personally identifiable information (PII)
- Information that can identify a person directly, such as a name or email, or indirectly when combined with other data.
- Pre-training
- The first stage of training a foundation model on very large, broad datasets to learn general knowledge and language.
- Preference data
- Records of which of several model responses a person judged better, often with reasons or rewrites.
- Protected health information (PHI)
- Individually identifiable health information held by HIPAA covered entities and their business associates.
- Pseudonymization
- Replacing identifiers with consistent codes so records can be linked to each other but not to a person without a separate key.
- Re-identification
- Linking de-identified data back to the people it describes, often by combining it with other data.
- RL environment
- A simulated or copied software setting where an AI agent can attempt tasks repeatedly and be scored, so it learns by trial and error.
- RLHF (reinforcement learning from human feedback)
- Training a model to produce responses people prefer, using human rankings or ratings of model outputs as the learning signal.
- Sublicensing
- When a licensee grants rights in licensed data to a further party, such as a customer or contractor.
- Success fee
- A fee paid to an intermediary only when a transaction completes, usually a share of the deal value.
- Supervised fine-tuning (SFT)
- Fine-tuning on examples of inputs paired with the desired outputs, often written or checked by people.
- Synthetic data
- Data generated by models or simulations rather than collected from real people or systems.