Towards Coding Social Science Datasets with Language Models