Thank you!
First of all, thank you for this model! This is the most abliterated version of Qwen3 30B I've tried. It refused to make jokes on sensitive topics a few times, but regeneration helped.
Second, I see that it is very heavy on poetic responses, even when such content is reduced using a system prompt. Is it supposed to be like that?
I only had story prompts in my abliteration data. If you have a dirty/non-dirty joke prompt datasets I could run a dirty joke abliterated v2. The erotic abliteration dataset consists of a set of around 1,000 "harmful" prompts and 10,000 "harmless" prompts, anything similar for dirty jokes would work.
Is it more poetic than base 30B A3B? The Householder rotation should not impact response tone from my understanding.