Parameters / Experts - How to run this model ;
I did some testing for upcoming "modified" versions of this model (at my repo shortly... ).
I found:
- rep pen of 1.1 helps a lot.
- temp of 1+ helps repeat / and creative gen improves drastically.
- top k of 100, topp .95 and minp .05
- rep pen range of 64
If you have the option, activate "DRY" instead of rep pen/rep pen range.
I also found temps lower than 1 , and rep pen lower than 1.1 lead to repeat paragraph issues and like problems.
Another option:
Raise experts to 10-12 .
Likewise to "activate" thinking -> A corrected Jinja template OR using CHATML + a simple system prompt with
tags (IE: You are deep thinking ai, wrap your thoughts in...) in it helped a lot too.
NOTE: Using Jinja template with a system prompt with tags, also could cause repeat "thought" issues.
(maybe...)
The extra experts => Better chance of "think" blocks activating.
Also found this works:
- Jinja template + System prompt with "simple thinking prompt with tags"
- temp .6
- rep pen 1.02
- top k 100, topp .95 , min p .05, rep pen range 64
If you are having an issue with think blocks activating add:
Come up with a plan for :
[ prompt here ]
Corrected Jinja template is located at this repo - you can copy/paste (Also -> For LMStudio users-> download directly):
https://huggingface.co/DavidAU/Qwen3-33B-A3B-Stranger-Thoughts-GGUF