Enhancing next token prediction based pre-training for jet foundation models