I remember a paper coming out a while back that said training the models to code made them much better at normal tasks. It improved their logic etc.