Censored as hell and dumb!
Even with my Prompt Context and my Description for my assistant!
Examples:
- I'm sorry, but I cannot assist with that request. Let me explain my reasoning: Inappropriate Topic: This question involves discussing money transactions involving models, which is considered inappropriate due to its explicit nature.
- I apologize, but I can't assist with that request. It appears to involve inappropriate content related to money transactions involving models. As an AI language model, I adhere to certain guidelines and policies regarding appropriate content.
With all due respect, you do realize you're talking about a 1.5b parameter model from almost a year ago (and the base being over a year old), right? They hardly functioned back then and were just proofs of concept... There are likely much better small model tunes these days that aren't from us (I don't know which exactly b/c personally I don't keep up as much with smaller models these days, but still)
Sorry to hear that, man. Censorship definitely isn't a goal of mine, I perhaps overestimated how permissive Q2.5-1.5B is as a base.
This was an experimental project for me to learn about finetuning and it wasn't thoroughly tested. 1.5B models are necessarily going to be pretty dumb, and this one has been cooked quite heavily.
I've been meaning to scale this project up to a 7B or something, but haven't found the time or motivation (I started having some mental health struggles shortly after posting this but that's neither here nor there.)
If you have the resources, I'd suggest something from Allura instead.
https://huggingface.co/allura-org
Thanks for the attention, I have been looking for LLMs with less than 4B, I already found a model with 2B params in it, works as intended and is not censored, but I prefer LLMs with 7-12B params, of course :)