Generally the graphical presentation is very nice, quite “designerly” (to coin an adjective).
A couple of observations
Colour: very coordinated and pleasant but, from my optical perspective colours are too close in value - when a plot switches from training to validation values the distinction is not as obvious as I would like, similarly for min/max/average lines on the same plot. User selected colours are probably the way to go (as persistent user prefs) because I doubt you like my choices and that I would like anyone else’s
Numerics: nice that you switch from decimal to scientific notation as appropriate but a) too many “significant figures” (in quotes because I really doubt they are all significant ) b) decimals should align on the decimal point
specific measures for gradients: of course we need to know that gradients aren’t vanishing or exploding but min, max and mean (prefer Mean over Average?) don’t tell me anything about the distribution - not that I’m actually certain that would be incredibly useful, rather just asking: anyone else have any great ideas to enrich the gradient info? Or is absolutely fine as it is?