Benchmarking code LLMs using user preference

This is neat. I feel like it should be blind by default though.