Recommended sampling params
#2
by
owao
- opened
Hey guys, what do you recommend? I see in your usage snippet you don't set anything and the generation_config.json is not setting any of them.
Should we use the transformers text-generation pipeline defaults? E.g. greedy decoding?
Thanks by advance for any answer, and thanks for releasing so many models the buzz is audible!
Only info I found was in your paper about temp 0.6 being the sweet spot. But what about top_p, top_k, min_p and repeat_penalty?