A quick idea.
I tought I try an Udacity course on machine learning. However the notMNIST_large contained over 200 000 files so it kind of broke my phone.
From this I got the idea to video compress the images to ten files corresponding to thousands of images from the letter A to J.
I think you can extract the images with some python module. The important thing is that this reduces 200 000 files down to 10. Which is much more friendly.
Another idea would be to quickly view the image set. Just play it as a video in mpv.
The size from the default settings got it down to just about half the size. To my surprice. But .png compression is pretty good.