[cypress66]

It's hard to think of a benchmark for these LLMs that wouldn't be massively overfit (I don't mean via training, but via hyperparameter tweaking).