Hyperparameters are the training settings humans choose rather than the model learns: learning rate, batch size, network depth, regularization strength, LoRA ra
A training setting humans choose rather than the model learns: learning rate, batch size, network depth, regularization strength, or LoRA rank. Hyperparameters shape how, and whether, training succeeds.
Parameters are the weights a model learns; hyperparameters are settings you choose before training; and inference settings like temperature are different again, though all three get casually called 'parameters', which causes endless confusion.