Förtuning

Förtuning, or fine-tuning, is a machine learning technique in which a model that has been pre-trained on a large dataset is further trained on a smaller, task-specific dataset to adapt it to a particular application. This approach is a core component of transfer learning and relies on the model's previously learned representations to improve performance with limited labeled data.

The typical process involves selecting a pre-trained base model, assembling task data, choosing a loss function, and configuring optimization with a learning rate typically lower than used during pretraining. It also includes deciding which parameters to update (all or a subset), and monitoring validation performance to avoid overfitting. Data preprocessing, regularization, and early stopping are common practices to improve robustness.

Variants of förtuning include full fine-tuning (updating all parameters), feature extraction (freezing early layers and training a classifier on top), and more modular approaches such as adapter modules, LoRA (low-rank adaptations), and prompt or prefix tuning, which keep most of the base model fixed and learn small additional components or prompts.

Applications and domains for förtuning extend across natural language processing, computer vision, and speech processing. Evaluation metrics depend on the task, with common examples including accuracy, F1 score, BLEU, and ROUGE. Strengths of the approach include data efficiency and the ability to leverage large pretrained models; limitations include the risk of overfitting on small datasets, catastrophic forgetting, data leakage, computational cost, and potential propagation of biases present in the pre-trained model. Practitioners choose strategies based on task requirements, data size, and available resources.