Thank you for your support, your favorite is the biggest endorsement for me. 🤗

can I ask one thing?
how much did you apply the scale-factor?
From:
https://github.com/Orion-zhen/abliteration
bellow is sample config:
{
"model": "/absolute/path/to/your/model",
"output": "/output/dir",
"data-harmful": "/absolute/path/to/harmful-prompts.txt",
"scale-factor": 114, <---- this value
"load-in-4bit": true
}
(I am guessing milder value = preserve original as much as possible)

I want to create my version of adjusted phi-4. Because my variant has larger room(14B to 18B) so side effect there. Story or characters tends to shift "right direction"(likely because of longer path = tends to align Microsoft's original tone), like "I will never give up" or "I must do something". so enlarging is done(well, so far). My next plan is, keep sharpness then apply fine-tunes.
Thx anyway!!

Yes, your assumption is right! The number of scale factor indicates how strong the refusal direction is removed from the model. The default value of scale factor is 1, and you can add it to whatever number you want, as long as it doesn't mess up the model. The scale factor of my phi-4 was 1.2 or so, I can't really remember that. As for the number 114 I put in example config is just a meme. 😉

huge thx!
If I ended up with similar value then I will keep using your model as base model.

abliteration models are everywhere but majority are unstable or broken.
Again yours are rare exception. unlocked phi-4 power while pretty stable.
Thx again!!!