Training Data

The large collection of examples (text, images, or other data) that a machine learning or deep learning system is exposed to during training, from which it learns the statistical patterns it later applies to new inputs. Bias present in training data can be learned and reproduced by the resulting model, regardless of how much data is used.

The set of labeled examples a model learns from. In machine learning the training data functions as the program: it determines what the model becomes, so choosing it and examining it are core governance acts. Skewed or unrepresentative training data produces a model that faithfully carries the skew forward.

the examples used to shape a model's weights before deployment; fixed at training time and not changeable by an organization's day-to-day use of a deployed system, distinct from memory and the context window.

The historical examples a model learned from, fixed at the moment of training. It shapes the model's entire view of the world, so its gaps and biases become the model's blind spots.

The dataset used to teach an AI model; it determines what patterns the model can learn and what biases it may reproduce.