Not at all uncensored
Ive been testing this for a while now and its very censored. Sorry but the training did not uncensor it. Dolphin did it much better
https://huggingface.co/cognitivecomputations/dolphin-2.9-llama3-8b
My guess is because the instruct variant is much harder to break censorship on
My guess is because the instruct variant is much harder to break censorship on
Why would the base model need uncensoring at all?
My guess is because the instruct variant is much harder to break censorship on
Why would the base model need uncensoring at all?
Possibly people are looking for a overtly toxic model? It's sometimes fun to talk to an unhinged model. Like some solar based models that openly admit to wanting to manipulate humans.
This is a solar model with a uncensored assistant card, It just inserts random crazy stuff into answers.
@rombodawg You don't need finetuning that much with Llama 3 - you can just play with template and system message a bit, and original model can be pretty fun then, too. They did it pretty well with Llama 3 - default template and system message makes it reasonably safe for general audience / kids (when model is in the assistant role), but it can also be pretty creative / playful with custom templates, when you want it to act as less boring characters. Dolphin is less censored by default, but finetuning also made it less smart (it started messing up basic math, etc.). Samples below using original weights from Meta, just different configurations (32k is just different rope_scaling configuration during conversion - weights themselves are the same):
I tried it. It's definitely not uncensored.
I see no big difference between this and the Meta AI chat. Fucking useless bot says it can't create explicit content even if there is no mention of explicit content anywhere.