this would be useful outside of coding. could you release a benchmark so we can have more models tuned for this?