edit: I found this:
https://scrollprize.org/data_browser#/samples/PHercParis4/se...
The JSON seems to suggest that I'm mostly looking at ink detection output, but I could easily be using the tool wrong.
But I also found this awesome explanation:
https://scrollprize.org/data_fragments
I guess I bunch of the training was done by using fragments of scrolls where ground truth data is available using IR photography.
Also... that xray resolution is absolutely amazing!