3 ms·
Are those numbers real in the technical report? If so it really is something big, a < 4 billion model overtaking SOTA 7-8 billion ones, and by quite a large mar
by margorczynski 2y ago
Are those numbers real in the technical report? If so it really is something big, a < 4 billion model overtaking SOTA 7-8 billion ones, and by quite a large margin.
Seems too good to be true, or I'm missing something?
- unraveller 2y agoOthers are very sceptical it reaches llama3 8B level and claiming some mistakes in their benchmark numbers[1] of phi-2 at very least. they are updating fixes to the Phi-3 models on a few things very quickly so you will be able test soon but so far in my testing of 4k mini it has some properties that are better than even llama3 70b, namely admitting faults and correcting it once pointed out instead of hiding them again more creatively. Whatever inner intent balancing act it does seems more desirable to me for code things. [1] https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/discussions/20 https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/di...