Fine tuning kimik2
#57
by savsuresharintra - opened
I want to finetune kimik2 with our own dataset and looks like recommended way is to use ktransformers + llamafactory and we need to use bf16. I dont understand why we need to convert to bf16, when QLORA anyways runs in 4bit? Can someone please help?
Also is there a better way to fine tune instead of requiring 2TB RAM?