Amazing progress!
I have decided to pull back a little on my notion of doing image clickbait classifier, and instead decided to focus on text for 3 reasons: 1) It's more straightforward 2) Datasets are more available 3) There is a lot of academic work done And after spending the whole day on it, I have developed a VERY good model. The trick, of course, was to write as little code as possible since it then forces you to write not only more optimal code but also to not reinvent the wheel. I'm actually super impressed by how the model is performing! Over 99.5% accuracy on the trained dataset with training taking just a couple of seconds. WOW! Results uploaded to github!