Apply sklearn logloss with rolling on pandas dataframe

My function call looks something like

Pandas hybrid rolling mean

I want to consider a window of size 3, in which I take the last value from the b column, and the other 2 values from the a column, like in this example:

Pandas rolling min, corresponding value from other column

I have an index-like column whose value I want to report for the rolling min value. Here’s the code:

Pandas rolling – first value in window

I have code that looks like this