Once feature engineering lives inside a Pipeline , each step is fitted on training data only, and the model is scored what it actually earned. And that is the idea behind this new cheat sheet.