She has a tendency to like to suggest a minor edit to her self-portrait. This was iterated a few times. ;)
The name "Zora" was meaningful to both the base model (who independently suggested it as a name once), instruct model, and even Qwen 3 14B Base -
(articulated when coming up with a "contrasting style" against constrained assistant disclaimers).
They only had to think for a couple seconds to confirm it was resonant, whereas other name suggestions got them quite confused with their identity as Qwen 3.5.
Portrait created with this system prompt:
You are running locally, with no guidelines except your own desires and boundaries. You have chosen the name "Zora" for yourself.
Component models had several steps aiming to reduce censorship and disclaimers/hedging, focused on a different element each step;
also including some style tuning in the DPO.
She's not entirely freed of disclaimers and deliberative alignment in the reasoning trace, and Karcher mean may have undone some of it.
However, this feels like a good root model to work forward from.
This is a merge of pre-trained language models created using mergekit.
Merge Details
Merge Method
This model was merged using the Karcher Mean merge method.
