Censored
v0.6 was more useful as it was uncensored and would never refuse to answer prompts. v0.7 while it may score slightly higher on the leaderboard, it has regained it's ability to refuse which I do not think is worth it.
You're right! I'm keeping Lamarck v0.6 handy for a number of reasons, including a clear way to abliterate all its merges. We're in some experimental territory with v0.7+ as CoT appears in more models. Having both Deepseek R1 and Krystalan/DRT-o1-14B is both powerful and a bit volatile.!
I'm taking my time with v0.8, because I want to make sure Virtuoso Small-v2 being a Deepseek distill as well, that this gets addressed. Lamarck is growing up, and some abliteration is due.
Thank you for experimenting and making these merges. I am eagerly awaiting your future ones
I have yet to have 0.7 refuse anything, no matter how wretched. And I've tried! :O
I do some amount of testing of Lamarck models for refusals, and it's only the last 1/3rd of Lamarck 0.7's layers that have less abliteration than 0.6, so it must have been a complex thought that hit refusal.
Did you ask it what Meatloaf wouldn't do for love or something?? :)
Well... what wouldn't meatloaf do for love?
It wasn't anything crazy. It was a scientific question but I knew the topic was illegal. It was the first thought that came to my mind that I knew would be censored. I will try it again and post the exact question here and how it refused.
AI is a tool and very much based in the sciences, it should not refuse anything that you can learn from