MiniMax M3.1 Flash resists abliteration, with no HF releases
The author says MiniMax M3.1 Flash is one of the harder models to abliterate, noting there are no abliterated releases on Hugging Face while other models have many.
I love that abliteration has gone mainstream, and the tools out there make it easier than ever to get started, but newer models are getting trickier. MiniMax M3.1 Flash is a good example, it's one of the harder ones I've worked with. There’s a reason there are no MM3.1Flash abliterated posts on HF whereas for every other model that Heretic can help with, there’s an abundance of “abliterated” model posts. I think going forward a lot of it comes down to understanding where a model's behaviors actually live and how it predicts its words, and that's the part I want to dig into. Abliteration is not only the art of uncensoring a model but comprehending how these next word predictors derive behaviors, and there’s so much room for this section of ML to be explored and utilized to advance LLM’s. waitlist for @LyberAI is live at http://lyber.ai
This post quotes another post: x.com/lyberai/status/2114807653356720292