Evening-Truth
This is an unofficial guide to Google products. I'm not affiliated with Google.
Gemma 4 31B Base Prompt WIP
Overview
Surprisingly good for its size! Follows instructions, not overly sycophantic, and uncensored as far as I can tell.
It's a decent LLM for roleplay.
It's tricky to get actual thinking tokens out of it. At least I didn't manage it on SillyTavern.
It's still WIP... I gotta try some more instructions on it.
Setup
Tested and prompted via DeepInfra and NanoGPT API.
Generation Settings:
- Temperature: 0.9 to 1 works fine. Still testing though.
- Max P (Top P): 0.95
- Top K: 40-100 (DON'T LOOK AT ME LIKE THAT! 😂 The explanation is ⬇)
- Penalties: (using -2 to +2 range)
- Keep your sweet hands off. Let it cook without interruptions.
Yes... I know Anoki the Top K hater tells you to set Top K. Before you ask, no, I'm not having a stroke.
A nice person from the community explained it to me:
"Minor nitpick: top k should be set at a high value like 40-100 so that samplers after it (namely top p and min p) don't have to process the entire size 250k vocabulary for each token, it should result in a slight speedup that you'd otherwise leave on the table. This doesn't affect creativity because sane top p or min p values will almost always leave you with less than 40 tokens anyway"
Thank you so much Velocita84 😘
SillyTavern
Prompt-post-processing: Strict or semi-strict