Pretraining Federated Text Models for Next Word Prediction