I build something similar with a basic classifier in Tensorflow: https://github.com/Savjee/yt-spam-classifier

Here's the training dataset that I've been using. It includes all comments from my channel from last year until now. And I've manually tagged about 3000 of them as spam. https://docs.google.com/spreadsheets/d/1QEQrLne1SDxwQVl5qpGQ...