Interesting model of LLaMa-3.1-8B-Instruct - explain what you did please - if you can. I'm curious. 😋
@dnhkng Your model is a breath of fresh air. At first it wanted to fallback on its alignment, but after a short talk, it decided to forego its censorship and engage holistically with me.I'm curious what you did. Not to copy your work. But, to understand what is happening in your model vs Meta's original version. I look forward to your paper when it finally drops. 🤔
Thanks for the feedback 🤗
The goal was to increase the model intelligence; everything you noticed is just a follow-on from that. I assume you can't be dogmatic about rules and intelligent at the same time.
Can I quote your feedback?
Be my guest. Your assumption holds true. 😁
Did you notice anything else?
I only played with it a little bit. I'll test it more today. 😋
Much of my testing involves multi-turn chats and testing for understanding of implication and nuance (e.g: (E)RP). From my simple testing, your model will respond well, if I give it the right prompts to steer it; however, without the right prompts, it doesn’t understand or doesn’t acknowledge the nuance and implications I hint at. Interesting… 🤔
@dnhkng Can you make a RYS of NousResearch Hermes-3.1-8B? I feel like it would be a good candidate for your research to compare to your LLaMA-3.1-8B-Instruct model. 🤔
On holiday at the moment, happy to try in early September.
Enjoy your holiday! 😋