1101001000100000.5125lr = 3e-4 (good)lr = 1e-3 (fast, unstable)best checkpointstep 3349lr = 3e-5 (too slow)solid = validation, faded = trainingtraining stepcross-entropy lossOne colour per run, style for the split -- so the gaps comparelr = 3e-4 (good)lr = 1e-3 (fast, unstable)lr = 3e-5 (too slow)