The 'samples-per-category' parameter used when training this vllm semantic router

[RixinLiu]

Hi, you all! I'm trying to figure out how much data you sampled from each category of the MMLU-Pro dataset when training this router. Is it the default setting of 5 in the mmlu_pro_vllm_eval.py script? Because I have not yet found the relevant settings in the code, please forgive me if I have overlooked something.

[rootfs]

@RixinLiu The bench params are now exposed in #54