Evaluating Large Language Models Trained on Code