Important note: the metrics starting with eval/bench_ were done on a small subset of actual benchmarks due to interest of time, with less than 5-10% of each benchmark used on average.