Continued pre-training (CPT) is a training technique that extends the pre-training phase of a foundation model by exposing it to additional unlabeled text from specific domains or corpora. Unlike supervised fine-tuning, which requires labeled input-output pairs, CPT trains on raw documents to help …