During the pre-training phase of multilingual language models, the model is trained on large-scale unlabeled text data from one or multiple languages to learn general language representations.