Since all prompts must be converted to the prompt template, how do people go about reproducing EleutherAI's lm harness evaluations listed in the blog post? Wondering if there's an easier systematic change than changing the test harness code directly.
