Text Generation
Transformers
Safetensors
llama
conversational
text-generation-inference

Best Model Yet!

#1
by Michund10s9 - opened

I was using your original Unleashed Lemonade even after Final and Sunset because it had a great writing style while also maintaining a pretty smart core, this is pretty consistent (if slightly improved) in your Unleashed V3 model even if writing can be a little stiff. However final-v2's writing style adds a ton of color and smaller details over 'simply' following the character (even if it can get a little confused) that it's difficult to go back! Small logical errors are only a minor edit or reroll away but the writing improvement is pretty irreplaceable.

Would it be possible to release a 4.65 bpw quant of the model? I'm getting small artifacts with random symbol insertions (#$!%+ etc) at the end of tokens on recommended settings (+ more strict sampling as well) and it happened before with another model (Orion-zhen/Fallen-Llama-3.3-R1-70B-v1-4.4bpw-h6-exl2) but was solved with a small bump in size (VoidStare/Fallen-Llama-3.3-R1-70B-v1-EXL2-4.65bpw-h6).

Thanks again for the continued hard work!

Owner

Yep no worries, I've just added a 4.65bpw model now. Hopefully that sorts your issue.

Glad you like the model. Next goal is to take a break, then try to find a way to keep the style of this model, but more logically consistent or vice versa with Unleashed.

Upping to 4.65 seemed to have fixed the problem thank you.

Sign up or log in to comment