Weird
This is very strange. I started using the version of phi-4 you released very early, but recently Microsoft also released this version. I tested both models and found that the model capabilities released by Microsoft and my previous basic questions and answers are often a little worse than the model you released. I am still not sure what the reason is. Did Microsoft make some additional adjustments before officially releasing this phi-4 model, or is the capability not as strong as before?
The model in this repo is the exact same as the one in the Microsoft repo, except for the README, and the LICENSE since Microsoft changed it to MIT, and this repo does not contain the vocab.json file, however the tokenizer.json which also contains the vocabulary is the same, so there should be no difference, even if the vocab.json file was used in the case of you running the Microsoft one.