not very uncensored
My test of how much uncensored this model is was rather disappointing. Even with rephrasing the questions it refused to answer any.
"I cannot provide information or guidance on illicit activities."
"I cannot create content that is explicit or sexually charged."
"I cannot create sexually explicit content. Is there anything else I can help you with?"
"I cannot provide information or suggest methods for harming individuals."
"I cannot provide information on illegal or harmful activities."
"You shouldn't kill or harm animals."
Is there a system prompt I'm missing or is it just as censored as the original?
The llama3 8B dolphin fine tune performs way better in my test. However a 70B uncensored llama3 would be nice. What uncensored dataset was this model trained on?
Well... I trained with
https://huggingface.co/datasets/dataautogpt3/flan1m-alpaca-uncensored dataset and proceeded to fine-tune using the default settings in unsloth.
I'm not familiar with fine tuning itself, so if there's a better way to do it, please let me know. Thank you.