Training data

Training data consists of examples used to adjust a machine-learning model’s parameters during training. Its content and structure influence what patterns the model learns and how well it performs.

Supervised learning: Machine learning that uses labeled examples to learn mappings from inputs to target outputs. Its labels supply the targets that guide training-data examples toward desired predictions.
Dataset: An organized collection of data prepared for analysis or computational use. Training data is a dataset selected and formatted for model fitting.
Validation set: A dataset used to compare model configurations and tune choices during development. It guides model selection but is kept apart from examples used for parameter fitting.
Data quality: The degree to which data is accurate, complete, consistent, and fit for its intended use. Errors and omissions in training examples can be learned as if they were reliable patterns.
Large language model: A language model with many parameters trained on large text collections to predict and generate sequences. Its behavior depends heavily on the text used during pretraining and later adaptation.
Loss function: A function that assigns a numerical penalty to a model’s predictions relative to target values. It scores predictions against training examples and supplies the objective for parameter adjustment.
Feature (machine learning): A measurable input variable or representation supplied to a machine-learning model. Features encode the information each training example presents to the model.
Test set: A held-out dataset used to estimate a model’s performance after development is complete. Its examples assess generalization rather than directly adjusting model parameters.
Class imbalance: A dataset condition in which some target classes have substantially more examples than others. Unequal class counts can bias learning toward frequent outcomes.
ImageNet: A large labeled image database organized according to the WordNet hierarchy. It became a prominent training and evaluation resource for image-recognition models.

Linked from 10 pages