Good, but it doesn't stop
Ciao Benjamin,
I agree with you that this 270M model does have really a huge potential.
Your ORP version Guanaco is certainly better at following instructions...
but is there any tricks to make the model stop generating?
It is behaving like a completion model (gpt2 style)...
Hello Fabio,
I fine-tuned the model with the default chat template.
But I cannot say that the model is good, or that it will stop to generate at the right time... I think it is better than the official instruct model released by Apple, but still extremely bad...
ahaha you are right. it is better. I would like to instruct fine tune too. But I don't even know from where to start. Any hints, maybe you have already written something about it?
Yes, I have written about it in my newsletter:
https://kaitchup.substack.com/p/fine-tune-tiny-chat-models-with-apple
I'm still considering whether to post something similar also on Medium.