Training data is the collection of examples used to teach a machine learning model. For a language model that means vast amounts of text; for an image model, huge sets of images. The model studies this data and absorbs its patterns, which is why the data's quality, breadth and biases end up baked into the model's behaviour.
This has real consequences. A model trained mostly on formal writing will sound formal; one trained on data with a particular bias can reproduce that bias. It also means models have a "knowledge cutoff" — they generally do not know about events after their training data was collected.
Related terms
Machine Learning (ML) · Neural Network · Large Language Model (LLM)