Continuing to train a pre-trained model on a smaller, task-specific dataset to adapt it to a new domain or use case with far less compute than training from scratch.