Sao10K

/

L3-8B-Niitama-v1

like 27

What settings are you using? Lunaris is still by far the best model I've experienced, would appreciate tips!

What settings are you using? Lunaris is still by far the best model I've experienced, would appreciate tips!

1.25 temp, 0.050 min p, 0.8 DRY multiplier (on the newest koboldcpp) + LLaMA 3 system + context formatting. The way you write your characters is also crucial. Sao's tunes do amazing once you've given the card 3-4 message examples that never get pushed out of context. Imo, Niitama is just as smart as Lunaris, but God... it's more soulful. Anyways, enough glazing, can't wait for the supposed long context L3 8B or even intermediate sized models(I dream of a 14B), so that we can get some better tunes out of them.