Dataset
A collection of examples used to train, evaluate, or ground a system.
A dataset is a body of examples assembled for a purpose: training a model, testing one, or grounding answers in real material. Its quality, coverage, and licensing usually matter more than its size.
A training set needs coverage of what will actually happen, including the awkward cases. An evaluation set needs to be held out from the training set, or the number measures memorisation instead of capability. A grounding corpus needs to be current and licensed for the use.
The most common mistake is treating a dataset as a thing you gather once. It is a thing you maintain: it drifts as the world changes, it carries rights that have to be tracked, and its labels get wrong in ways nobody notices until the model reproduces them.
Related terms
Training
The expensive process that sets a model's parameters by exposing it to data.
Evaluation
Measuring whether the system does the job, on a fixed set of cases, repeatedly.
Data Labeling
Attaching the right answer to each example so a model can learn from it.
Data Governance
The rules and habits that decide what data exists, who may use it, and how long it lives.
IP
Who owns what the model produced, and what it was trained on.
Want this working inside your team?
It starts with one conversation about your workflow.