Awesome. I Got Very Good Responses, However...
This is by far the best performing fine-tune I've ever used, not just among Mixtrals.
Other than excessive alignment it's a solid performer across the board. Months later and no community fine-tune of Mixtral comes close, which is odd because every other official instruct version not only performs worse, but MUCH worse, compared to community fine-tunes, including Yi, Llama 2, Mistral and Gemma.
My only complaint is that the alignment is often nonsensical. For example, I can ask for a joke about a man, but if I ask the same joke (word for word), but replace man with woman it refuses and lectures about being inclusive and non-offensive.
Alignment is VERY crippling and should be reserved for mitigating illegal or amoral actions (e.g. stealing a car or sharing celebrity info hacked from their phones). Using alignment to avoid potentially offending someone in a world filled with hypersensitive neurotics is neither reasonable nor possible.
Additionally, this LLM compulsively uses fabrications to enact alignment. For example, Milla Jovovich appeared topless in at least 11 theatrically released films (all consented to). Yet you fine-tuned this LLM to not only fabricate denials (no nude scenes), but also fabricate substantiations out of thin air (because she focused on being an action hero instead), and then fabricate explanations when called out (body double, camera angle, body paint...).
Ironically, this is having the exact opposite effect. When parents ask your AI model if there's nudity in films because they don't want their kids to watch it if there is, and you're fabricating lengthy denials about how actresses never due nude scenes because they're above them, and are instead focused on (insert fabrication here) you're causing parents to expose their kids to what they don't want them to see. I'm not singling you out. EVERY official fine-tune, even GPT4, does this, including Gemma, Yi, Qwen and LLama 2. And it's annoying because I'm trying to get my family interested in AI and they're sending back emails filled with such deliberate fabrications, refusal to output non-Disney friendly responses...
Anyways, with future fine-tunes please consider just telling the truth unless it's illegal or unethical. Not everyone is a young child living in the Disney universe. But still, kudos for still being, by far, the best fine-tune of Mixtral.