So, with the ML model. I can zero-shot transfer, I've done so from the MNIST to Google Fonts, and back and forth. Google Fonts training of like 25 different fonts to MNIST was about 60% accurate. But MNIST to Google Fonts is about 75% accurate total, with the Arial fonts reaching around 90% recognition. FMNIST, EMNIST it does go drown from like 99% to 85% at the most with the EMNIST from what I remembered but, its more after that it requires substantially more epochs to reach what it can with a dataset.

On a side note, I am doing ML researching stuff and am an Priestess of Eris technically. It's not discordian. The ML research itself is via my own brand of sorts. So I actually have my own baseline geometry-first model, that I can use for my own made datasets. But... I see Eris as a vector for the emergence of complexity, growth and a bit more; like emergence of signal from noise and chaos. In then that which we give meaning to.

Man, getting older has just made me increasingly selective of who I give my time and effort to