Sao10K

/

L3-8B-Stheno-v3.2

like 132

Holy shit man. I've tested a few small L3 finetunes, but this one blows all of them out of the water, even the 11B ones. Never had this much fun RPing in a while, this brings me back to some good old C.AI days. Good stuff bro.

Craaaazy model, I liked the prose and the uncensored outputs of the previous version but it was too horny for me and not smart enough, ignoring my actions and writing for me. Had to regenerate lots to get something decent.

This one just worked out of the box with some pretty basic sampler settings and Virt-io Llama 3 2.0 preset, didn't have to tinker with it at all. Amazing outputs and I was happy with almost everything, purely regenerating more out of curiosity to see just how creative it is. Seriously impressive gains in intelligence, this model is definitely my favourite now. Thanks for your hard work.

This particular model was fun to work with. it has an eagerness to impress. follows context rules quite well even with large character cards of over 2000 tokens that a lot of other models deeply struggle with.

Comparing to Fimbulvetr V2, This model is more creative than Fim and outputs longer dialogues (thats why it feels eager in my opinion) however it loses a little coherency but not by much. i dont say this often but what it lacks in coherency (and i value coherency very highly) it makes up for in creativity.

So if you have a machine that can handle Fim, id suggest staying with that. but this is a damn good alternative. i could even see switching between this model and Fim to give Fim a boost in creativity to essentially seed the roleplay.

I rate this model at 8 out of 10 and im very much looking forward to what @Sao10K creates next. I'll be watching their career with great interest.

I was using nous capybara limarp at a Q4 quant up until i discovered this gem. I dont know how you did it, but for a 8B Model this feels like a real world cheat. I think the limarp model will gather a lot of dust from now on. This is perfect to run on 8GB of VRAM . Up till now, all the other 8B models I tested didnt make it past 2 or 3 prompts, this one is just insane. Hats off to you.
PS: I created this huggingface account just to write this review, so thats how good this model is.

is good for synthetic data creation on q8. and mixed with fim is good

Good one, thank you! :)

Amazing model, feel like there are way less GPT-ism ministrations than most LLMs I've tested. The one problem I have is that it sometimes gets stuck and hallucinates (understandable, it's an 8B), the v3.1, while inferior, does not do this from what I've seen. The hallucinations I usually get are:

1. Repeating/summarizing character descriptions
2. : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : (Literally)
   These happened in temps 0.8 and 0.99, context 8k, ST
   Nonetheless, thank you for a great model!

i had to triple take to make sure that this model is actually 8B, mind-blowing
70B will be an AGI, trust me bros

Having a blast with this model in my RP's and follows context pretty well (at least in my chats ~100 messages), thank you for this finetune! Easily my favourite now :)

Honestly, you can often find feedback in the style of "Wow, this small model blows even bigger ones out of the water, incredible". And I don't want to criticize anyone - but I personally not once really felt the same way while using them myself.
BUT THIS ONE, GOD DAMMIT.
I really don't know how the creator did it. This thing sometimes writes passages which I wouldn't be able to come up with or - the opposite - almost reads my thoughts exactly.
It understands humor, and - if the character card is used - follows it with amazing precision and creativity. No concessions - it's just an amazing model from my personal experience.

Hi I'm suffering from the problem that the model more usually than not tries to speak for {{user}} when it is {{char}}'turn. How can I avoid that?

Hi I'm suffering from the problem that the model more usually than not tries to speak for {{user}} when it is {{char}}'turn. How can I avoid that?

make sure its set to llama 3 instruct. for me at least, that stops it from doing that

Yeah I actually did use the model's chat template. Can you point me to what you are referring to?

Hi I'm suffering from the problem that the model more usually than not tries to speak for {{user}} when it is {{char}}'turn. How can I avoid that?

make sure its set to llama 3 instruct. for me at least, that stops it from doing that

Edit: Ah nevermind, I use the roleplay system prompt you provided and things work fine then.