Pretraining is the process of training a model (often on a large, generic dataset) before fine-tuning it on a target task. It helps the model learn rich, generalizable features, reducing the amount of domain-specific data needed. Pretrained models like BERT or GPT have revolutionized NLP.