Dang. I guess LLM folks don't usually audit and QA their training sets? This one just trains the model to output invalid json all over the place 💀