Got Excellent Results in A RP Using This Model While Neutralizing Samplers

#1
by AlkaliV2 - opened

I had used other models in the past and they all seemed to get into sort of a repetitive style response loop. But between the info here about neutralizing samplers: https://www.reddit.com/r/LocalLLaMA/comments/1d5fyhb/llama_3_repetitive_despite_high_temps_turn_off/

And the Llama 3 context/instructs you posted I am getting a very intelligent and unique experience from a non-complicated card. I'm also using an AMD RX6800/RoCM so I'm no stranger to unique configs. I did the following using Ooba and SillyTavern:

  1. Downloaded, imported and set your Context and Instruct presets from this post into SillyTavern
  2. Imported Sampler settings Lewdicu-Samplers as a starting point (https://huggingface.co/Lewdiculous/Model-Requests/blob/main/data/presets/cope-llama-3-0.1/3.1.0-Lewdicu-Samplers.json)
  3. In Silly Tavern > Presets clicked the button "Neutralize Samplers", Set Context to 8192 and Response to 512
  4. In Silly Tavern under Advanced Formatting > System Prompt I entered "You are {{char}} and fictional character in a never-ending roleplay with {{user}}."
  5. Used a 700ish token character card with no special formatting and ran a test

Not only did it have full understanding of the card, it actually had personality. It didn't rush the story, it answered mid-scene questions not included in the character card, responded in clever ways, it didn't produce any blank killing the chat and was incredibly detailed. Quite impressive and it fits in my 16GB of VRAM with 8192 context so I am very happy. Thank you for making this model, it is very impressive and uncensored!

Edit: Forgot to mention, I am not on the dev branch of Ooba so while the Reddit post above talks about DRY I did not use any of it in my settings even though it was present in Silly Tavern Presets.

Hey, thank you for your feedback!

Sign up or log in to comment