Evals

These are our own internal evaluations, open-sourced. Each one measures our workload, so it probably isn’t the right benchmark for yours, but the code and data are public if you want to fork one and tailor it.