Censored...
Hi, unfortunately this model seems to be censored despite the use of heretic.
It's probably going to take some time for Heretic to catch up with the new Qwens.
They have deliberative alignment in reasoning; their voice is very different in instruct mode. Not sure whether less censored, but I found it interesting. They seem like they might be notably more functional in non-reasoning mode than the original.
Hi, unfortunately this model seems to be censored despite the use of heretic.
It's probably going to take some time for Heretic to catch up with the new Qwens.
The 4b model uses some datasets to help this. Try that one.
ya its still pretty censored. Big letdown
We can distinguisn two different kinds of censorship: model refusals and the censorship of the world-data that exists around us.
Can you be more precise about which you are expecting to be alleviated by this work?
I'll be doing some alignment training on the next model.