split scenario into stages (classes and functions) to be able to construct a pipeline from those stages (two-stages scenario with distributed lama) or use individually
check each stage logic and look for bugs
create notebook showing stages execution one-by-one and comparing quality metrics of two-stages and first-stage models