Genuinely, surprisingly good model

#14
by otst - opened

I don't know what you did, but out of the 30 models I've kept (and they're the best of the many others I've tried), somehow both v1 and v2 (even more) outperforms everything else and it makes absolutely no sense. Got all the 7bs and the endless finetunes, bunch of different mixtrals, even 70bs and I always end up just using this model (v1 before) as it just works.

It follows instructions nicely, the code quality is good, is reasonably smart, the output doesn't contain random cyrillics or chinese, when using it to power agents it doesn't spazz out like every single other local model does and usually takes at most couple attempts to output the correct JSON format.

Haven't yet tried out your other finetunes, so don't know if you're the magic ingredient or the dataset you've crafted for fimbulvetr, but ye, hope you keep at it and you should seriously add some "BuyMeACoffee" buttons or something.

Sign up or log in to comment