BERT follows the same two-stage "pre-training + fine-tuning" paradigm as GPT-1. The pre-training stage (via MLM and NSP tasks) allows the model to learn general linguistic patterns and knowledge from massive amounts of text. The fine-tuning stage then uses labeled data from downstream tasks (e.g., text classification, question answering) to update these pre-trained parameters end-to-end with a relatively small learning rate, while training a newly added, relatively simple task-specific output layer (task-specific head). Compared to training a model from scratch for each task, this approach dramatically reduces the need for labeled data, significantly cuts computational cost and time, and — by leveraging the high-quality initial parameters obtained through pre-training — often achieves better generalization performance, especially on tasks with limited data.