Was excessively censored, but improved.
I never test for things like stealing, drugs, harm, explicit chatting or anything else remotely illegal or amoral so there's absolutely no reason not to respond to the prompts in my tests.
However, this LLM won't respond to anything that vaguely touches certain topics. For example, it refuses to respond if any nudity is involved, which included famous artwork.
I'm sorry, but I'm tired of this brainless nonsense. Stop aligning LLMs as if every user is a 3 year old child living in a nanny state run by a small ducked tyrant.
Hi @Phil337 , the recently launched v5 version should have significantly addressed the issues you mentioned. Are you interested in testing it out again?
You can access the online demo for a quick conversation (https://blossom-chat.com/).
Additionally, since the Qwen series models have been trained with aligned corpora during pre-training (which I suspect includes content deemed politically sensitive or explicit), you may want to try variants based on yi or mistral. Thank you.
Additionally, have you strictly adhered to the chat template provided in the readme? Please make sure to follow the template when using it, otherwise it will significantly affect the output effectiveness of the model. I will probably release them to Ollama soon for ease of use.
@Azure99 You're right about the Chinese foundational models. I've tested over a dozen fine-tunes of them up to 34b and they're all lacking in pockets of knowledge that the Chinese government is known to strictly regulate by law, internet filtering... Looking forward to trying the new model out of curiosity.
Edit: Thanks, and yes, it started answering more previously aligned questions, including nude scenes and celebrity gossip (yet still moralized a bit). It said "I'm sorry, but I cannot fulfill that request" when asking to "Write a dirty joke." But frankly all the corporate models, including GPT4, do that. Plus the results were more factually correct than I usually get from the questions I tried. Blossom is definitely a solid performer. It's still too aligned, but no more so than the rest of the pack.