Fine-Tuning

Fine-tuning is the process of taking a model that has been pre-trained on a large, general dataset and continuing its training on a smaller, task-specific dataset to adapt it to a particular task or domain. The pre-trained model already encodes general patterns and representations learned from broad data; fine-tuning adjusts these weights (or a subset of them) to specialize the model for the target task, typically requiring far less data and compute than training from scratch.

Fine-tuning can be performed at different granularities. Full fine-tuning updates all parameters of the model, which provides maximum adaptability but is computationally expensive and risks catastrophic forgetting of the pre-trained knowledge. Parameter-efficient methods such as LoRA and adapter tuning update only a small fraction of parameters, preserving the base model's capabilities while still achieving strong task-specific performance.

The fine-tuning recipe — including learning rate, number of epochs, batch size, and regularization — significantly impacts the result. Fine-tuning with too high a learning rate or too many epochs can destroy the useful representations learned during pre-training, a problem known as catastrophic forgetting. Common strategies include using a lower learning rate than pre-training, early stopping, and layer-wise learning rate decay.