Early test version of what will soon be Indexnusrefather/Nyx-RP-9B-Instruct-2608-v2, for now its one of my first attempts with it and it might be quite unstable, check it out.
PS: This readme sucks, I know, it aint fancy.
Anyways, so the big difference from Nyx v1 is that this one had larger dataset, and I also trained it the different way because Nyx v1 approach turned out to break this one...
Here is my expression cleaning the dataset from Gemma slop btw:
5211162779375576356
So, uhhm, that is only rank 128, I will release beta(rank 160) and gamma(rank 200) later for comparison, there are safetensors and some quant(s) in the repo. And also I will now release the epsilon, the main thing about it is that it had the highest rank of them all, rank 216, and I also additionally filtered dataset for it, soon will be releasing it.