microsoft

/

phi-4

like 2.2k

Follow

Microsoft 17.4k

Hello creators of Phi4,

was phi4 trained with conventional fill in the middle format? i.e., the input to the model is

<|fim_prefix|> {CODE_BEFORE_CURSOR} <|fim_suffix|> {CODE_AFTER_CURSOR} <|fim_middle|>

and the output from the model is to predict the middle part?
I am asking because I tried this format but the quality of the output was quite disappointing compared to other models, so wondering if the format is different.