That’s a great idea! I suppose you’re thinking of secondary school “kaggle”! So… Welcome - and thanks for starting a great conversation
Just off the top of my head I can think of a number of things that would be quite straightforward - and have a competitive element through e.g. some metric such as Accuracy/Loss, or hitting a certain target Accuracy/Loss with as few neurons as possible…
For example, recognising MNIST digits is classic and straightforward, and would provide a good pedagogical introduction to certain concepts such as the forward pass, back-propagation, convolution - without needing to get into heavy/math detail (unless to stretch more interested students).
I think the way that PL shows what is going on in each component and during training would be great for students - pure code is not as engaging.
Let me think about some specific examples/datasets and I’ll reply again later, but in the meantime I’m sure one of the PerceptiLabs guys will have some great input for you.
You could also consider some of the many Live Coding examples that @robertl has shown on the PerceptiLabs YouTube channel… or have you already considered those and are looking for something else?