Very Nice. I Hope Someone Tries Fine-Tuning Phi-2 Like This.
I like the response style of this LLM more than any other I've tried.
They're short and to the point, but unlike others with succinct responses, this LLM gives longer responses when the circumstances or wording of the prompt call for it.
I can't find a usable Phi-2 because people are trying to make it too wordy (e.g. orange2), resulting in long strings of hallucinations, weird tangents, gratuitous moralizing... But without fine-tuning with a diverse dataset Phi-2 struggles to perform basic tasks.
Hi @Phil337
Thanks for the feedback :)
As you mentioned, the model tends to provide longer responses compared to other LLMs available. I ensured this by filtering datasets to generate longer responses and by creating multiple-choice datasets using Mixtral (rest assured, I provided the correct answers to it, don't worry :D) to explain the answers in depth rather than just providing them.
Regarding the phi-2 request, I may actually do it. I'll tag you if I do :)
Hi @Phil337
The Phi-2 version of the Einstein v4 is on the way. I hope it turns out to be something good!
It will be in this repo when it's finished: https://huggingface.co/Weyaxi/Einstein-v4-phi2
