Framed properly, it does detect what you posted: https://imgur.com/a/0wdqPea. Positively identifying abstract geometric loss while rejecting non-loss text and drawings was something I had to explicitly optimize the training sets for.