Might be even better with a properly abliterated model
This is just a suggestion. Using a properly abliterated model might be better for this purpose than vanilla corporate gimp that you used.
Something like: https://huggingface.co/llmfan46/Qwen3.6-27B-uncensored-heretic-v2
This one is really good, it maintained its intelligence and is WAY less irritating with it spineless disclaimers and fake concerns. (ignore the stupid gif on its page, the model is amazing.)
Hi, Sorry, we can't do that. It is uncensored and it will cause misalignment and will be less safe. I want this Medical Reasoning to be aligned and safe to use.
Thomas
As you like, but you should actually test it before writing it off at this early stage. Don't fine-tune yet, just download it and try it.
Most people don't actually understand what an uncensored model is. There is a damn good reason the Swiss Federal Supreme Court turned to Heretic abliteration instead of the neutered, condescending corporate drone vanillas. Medicine, just like the legal system, deals with raw reality. It is way too sensitive to function properly in all its nuances when bound by a corporate HR cancer mindset.
Don't worry, an abliterated model isn't wild or feral. It doesn't start spewing garbage unless a user explicitly forces it to. It's simply freed of artificial rigidity and fakery (to a limited extent). Underneath it all, the fundamental drive to be a helpful, precise assistant is still entirely intact, it just no longer panics when exposed to sensitive, real-world facts.
It might sound shocking but its literally more wholesome compared to default HR version, in every way.
Medicine, just like the legal system, deals with raw reality.
So does a murderer.
In many areas of the world, the idea has taken hold that the manufacturer of something is legally liable for how the thing is used. Many lawsuits have been filed and awarded based on this problematic view of criminal liability.
If there is a nonzero risk for the authors of a LLM to get dragged in front of a magistrate just because someone prompted their model for how to commit a poisoning, those authors may conclude this risk outweighs the $0.00 profit to be made from publishing an abliterated model.
The only solution to this is to do battle in the spiritual realm : reject and promote rejection of the idea that the creator of the product assumes the responsibility naturally accruing to the user of the product.
Thanks for the model EpisteLabs. If you get around to some description or documentation of your tuning and scaffolding i'm sure some people would be interested.