GGUF
Not-For-All-Audiences
nsfw
Inference Endpoints

I've got another (dumb? genius?) merge request for ya man!

#1
by BlueNipples - opened

So I found frostwind (of solar instruct origin) to be pretty darn smart (surprisingly so actually probs better than any <20b I've tried), but with terribly dry prose (worse than gpt). Understood everything but as creative as a rock. Nous hermes new solar instruct fine tune is MUCH better, but probably a little less coherent than frostwind. But defo better prose.

So here's the idea - I was hoping there was a way you could mash Noromaid 7b into some 10.7b frankenmerge (something I've never been able to work out), maybe with toppy 7b? Or whatever you think would be fun or clever.

And then combine the resulting 10.7b model into the other two in some clever way, such that the frankenmerge incoherency gets evened out by the normal models, but it inherits some more prose. A fine tune on something if you feel inclined.

I totally lament the smaller local models lagging so far behind the ones that need juicier gpu's rn. I know there's gold in here somewhere, and there might be enough fine tunes now to start to play.

Or not, and tell me to go eat dirt. As usual! I can't expect anything really you make SO many merges.

I should mention I partly stole the idea from this model:

https://huggingface.co/NyxKrage/FrostMaid-10.7B-TESTING-GGUF

Which I've tried and it is in fact pretty good. Best prose of any solar instruct I've seen, even if it's a little lacking in logic/coherency, and honestly probably just as good prose wise as some of the mistral small finetunes (I tried in the same rp using noromaid 8x7b, and ended up favoring this more often, despite it's frequent confusion it just straight up described things better). And probs more coherent than noromaid 20b, even still, tbh.

Having tried this model out now, I'm convinced there is something clever in this general direction. Perhaps the slight influence of the medical model also help the prose? Not sure, but I think they do sometimes do that. Suspect there's a merge or finetune in here that can beat anything current 20B and under.

Sign up or log in to comment