I have done some Deep Reinforcement Learning so I can tell you that making a NN is not the way forward here and it is far from easy. If you manage to take the NN of LC0 as it is and train it with another algorithm (according to how you want it), you can get your training software. Even designing this algorithm would be difficult.
The point here is that LC0 NN is capable of having a strength matching 3900 levels. It learns patterns in chess very efficiently. If you want a strength of 1600 or 2000 or whatever your rating is, it's definitely more than capable. Maybe you would need a new NN if you want a strength of 4500. Intelligence is already out there. You just have to customize it for your needs.
A much more practical thing would be to try to beat LC0 snapshots of past close to your rating. They should be available on lc0 site, you'll have to make the appropriate frontend. But you can get an easier version by trying to beat stockfish levels on lichess. Maybe you can request the lichess community to add more levels at smaller intervals or make some more yourself.
Earlier engines like Fritz and I guess the very old Deep Blue itself were trained on GM games so this is not anything new. Self play is much stronger and you'll be able to feel playing many levels how NN itself learns patterns step by step.
P.S. Such a thing would help Carlsen, Anand and others much more as you can learn through books but they can't. They can try to beat 2900 LC0 but you can learn from books. Everytime they crush the opponents who write the books we read.