teknium

/

OpenHermes-2.5-Mistral-7B

like 785

What sampling parameters are you using? I've found Mistral tends to hallucinate much more than other models at typical temperature settings. I use 0.5 and it performs far more accurately.

I moved this discussion to closed because the "intelligence" Hermes 2.5 gained across the board was well worth the slight increase in hallucinations at the fringes of Mistral's knowledge and marginally low TruthfulQA score (52). Thanks for your work. This is now my favorite LLM.