The new MMLU King

#3
by weezywitasneezy - opened

Congratulations on blowing away all the other models on MMLU by nearly 5 points! Most impressive! Are there any particular datasets, optimizations or tuning strategies that you think contributed to this exceptional performance? Thanks!

weezywitasneezy changed discussion status to closed
weezywitasneezy changed discussion status to open
NeverSleep org

Causal LM 34B Beta did already pretty well on MMLU but it was taken down, I trained on my usual RP datadet
It was not flagged, but I can't be sure it wasn't contaminated, I also wanted to do a retrain because one of my dataset had broken asterisk for RP but it was taken down before sadly

So yeah, big MMLU but take it with a grain of salt.

NeverSleep org

recovered now
see: https://huggingface.co/CausalLM/34b-beta/discussions/5

Thank you for making it available again!

Sign up or log in to comment