crestf411

/

L3.1-nemotron-sunfall-v0.7.0

like 12

Thanks for the feedback! It should work quite well for writing, so I’m glad to hear it’s working out.

It's not so much the prose quality. It's absolutely basic things that it gets right, such as understanding that instructions given by the writer are not known to the people in the story, and not confusing pronouns or people when there are more than two around, something chat models have trouble with, including your previous sunfall and story writer models. I didn't think it was possible with L3.1.

I suspect that my way of using models is totally unlike how other people use it, even for story writing.

But whatever it is, this model rocks, for me. At last, an alternative :)

Very cool! I’ve not had much luck the last 6+ months in general so the encouraging words are welcome. Thanks!

Well, at least I am always looking forward to a new crestf411 model every time I see one pop up in my list. And people often don't give positive feedback (nothing to complain about).

Maybe your lack of luck is due to lack of good and easy to fine tune base models recently?

Thanks for the kind words again! Yes I absolutely believe you’re right. I also think other model makers are incredibly strong right now and competition is harder than ever. A good thing, esp for users!

I have played around with this one quite a bit in recent days. Shivering spines are strong (i.e. lots of slop). More importantly, it destroyed all my stories mid-writing because I switched to it, and now suddenly the model follows all my instructions regarding style, language and more, and I suddenly have to tune down things to a reasonable level to get... a reasonable result. It's almost uncanny how I suddenly realise there is some unwanted content because... it actually followed my instructions, while previous models simply ignored most of them, most of the time.

It's so good that I feel I have to completely start over :)

Maybe it's to be expected with a L3.1-based model - being better than miqu - it's just that I tried a lot of l3.1 (and also miqu) finetunes, and they were all worse. Before, I would usually use QuartetAnemoi, sometimes mix things up with other miqu-based models or other models, just to return because they can't seem to follow simple prompts.

So... it's not perfect, but it totally restored my faith in llms (for story writing :) again, and is currently the top and only model I use for that.

Are you making any of this public anywhere? I would love to see what you're doing.

I do publish my stories, but for my sanity and my career or so, I keep it completely separate from this account, and hope nobody will ever make the connection... :)

So... I did fairly much with this model, and it's still my default go-to model. But I found that it's kind of crap (for me) when you use it with llama3 prompting. With llama3 prompt format it makes an astonishing number of logical errors and fails to follow my "simple" prompts. It took me a while to realise this, and this model is not the only one where this happens, but it is pretty dramatic, not only ignoring my instructions, but also becoming extremely repetitive.

I think the reason is that my way of using it to write stories is so unlike anything it knows that it works best when I use it if it were a non-instruct base model, i.e. as a text completer, although I use it in instruct mode with user/model pairs. It works stellar when I use some made-up prompt format such as "\ninstructions:\n" and "\nnarrator:\n" or some variant of alpaca, I think because I basically force it out of the normal assistant/chat mode.

Very strange. But after many megabytes of generated stories, I think it's not a fluke.

Anyways, thanks again for your work - I've since experimented with some L3.3 based models, and while they also work quite well, this one is still behaving better.