Tuning Language Models for Robust Prediction of Diverse User Behaviors