mradermacher

/

Poppy_Porpoise-0.72-L3-8B-i1-GGUF

like 3

No, but groups_merged.txt is about 50% of the imatrix training set (see https://huggingface.co/mradermacher/model_requests specifically "What is the imatrix training data you use, can I have a copy?").

I generally avoid quantising nitral-ai's models (unless I am asked to), simply because you already do an excellent job :) I write a bit more about that in "I miss model XXX".

You put together an awesome documentation. I might have to do something similar. I just assume people will figure things out.

My training data consists of about 160k tokens, about half of which is semi-random tokens (sentence fragments) taken from stories, the other half is kalomaze's groups_merged.txt and a few other things. I have a half and a quarter set for too big or too stubborn models.

Thanks for the clarification. Sounds good.

I wasn't rushing for this one since it wasn't a full version improvement and hopefully we should have the new Ortho models from them soon, which is the real hype for me.

Can't speak for anybody else, but I am super happy with your presentation, and didn't miss any documentation yet.

You put together an awesome documentation. I might have to do something similar. I just assume people will figure things out.

My training data consists of about 160k tokens, about half of which is semi-random tokens (sentence fragments) taken from stories, the other half is kalomaze's groups_merged.txt and a few other things. I have a half and a quarter set for too big or too stubborn models.

Thanks for the clarification. Sounds good.

I wasn't rushing for this one since it wasn't a full version improvement and hopefully we should have the new Ortho models from them soon, which is the real hype for me.

Unfortunately ortho hasn't seemed to improved the models any so far, so we may be switching gears here to get the opus finetuned version out instead while we still figure out if ortho will do us any good.

Can't speak for anybody else, but I am super happy with your presentation, and didn't miss any documentation yet.

I quite enjoy his work as well, much more effort into presentation then im willing to put into the models. (Although i've been trying to be better about it.)

No, but groups_merged.txt is about 50% of the imatrix training set (see https://huggingface.co/mradermacher/model_requests specifically "What is the imatrix training data you use, can I have a copy?").

I generally avoid quantising nitral-ai's models (unless I am asked to), simply because you already do an excellent job :) I write a bit more about that in "I miss model XXX".

Personally, I'm open to anyone quanting our models as it means the community as a hole just has more access to them. Thank you both very much for the work you do!