We train on a hybrid dataset of coding problems (based on MBPP), where each problem comes with a worked-out reasoning trace.