A hyperparameter is a configuration setting chosen before training begins, rather than a value the model learns on its own from data the way it learns its internal weights. Learning rate, batch size, the number of layers in a network, and how many training epochs to run are all hyperparameters: they shape how the model learns, but they aren't themselves learned.
Getting hyperparameters right matters enormously for how well and how efficiently a model trains. A learning rate set too high can cause training to become unstable and never converge on a good solution, while one set too low can make training prohibitively slow. Because the right values depend on the specific model, dataset, and task, teams typically search across a range of hyperparameter combinations, often using automated tuning methods, to find a configuration that trains effectively without wasting compute.
