What does Training Data mean?

Training data are the examples a model uses to derive its rules. In supervised learning, each example consists of an input and its correct outcome — a document and its category, for instance. Training data differ from test data in that the model never sees the test data until it is measured against it afterward.

It starts with selection: which cases are included, from which time period, and how the categories are distributed. Next comes annotation — assigning the correct outcome, done by subject-matter experts, usually with a second review for disputed cases. Preparation and cleaning remove duplicates and faulty entries. A common split is roughly 70 percent training, 15 percent validation, and 15 percent test.

The question of training data comes up for every custom model and every adaptation of an existing one. This affects text and image classifiers, forecasting models, and fine-tuned language models. Anyone using a finished model without adaptation, by contrast, is working with the provider’s data.

The effort annotation takes is regularly underestimated, but it pays off compared to simply gathering more data. A small, carefully reviewed dataset delivers better results than a large one with inconsistent categories. Because the labeling comes from people, it also stays possible to trace what a decision rule is based on.

Training data determine which cases a model even knows about. If certain groups or time periods are missing, a bias creeps into the result. That is why origin, time period, and composition must be documented — for high-risk systems, the EU AI Act requires this as well.