training: 200 steps final run, lr=1e-5

training: add early-stop callback + lr=1e-5 for next run