Grading does not work well for our model so we stopped doing it. Different teachers grade homework at different levels, and we have many many different teachers coming in and out. Often grading falls away after initial enthusiasm, so the "data" becomes sketchy and unreliable. It also places trainees into a child role, motivated externally by reward, instead of being in charge of their own learning. We do lots of 121 code review instead of grading, and focus on building real software together that works in the world, evaluated with good unit testing.