Inspired by Erik Hoffstad's Internet Comment Etiquette.
It's a simple RNN model built in Tensorflow and Keras, containing three layers: an Embedding layer, a GRU layer, and a hidden (dense) layer. It's trained on datasets I make myself using the Youtube V3 API for comments. The whole thing is put in some funky Flask framework, and tied together with some glue and sticks at the end.
Literally just do python app.py
As for currently model training yourself, start by moving to the youtube-api-parsing subdirectory (cd youtube-api-parsing) and changing the videoId to the video id of your choice (the string at the end of a YouTube URL). This will create a file that looks like {your video id}.txt in the appropriate folder.
After that, return to the main directory (cd ..') and go to models.py and change the videoId to the same one used earlier and then run python models.py. You can always change the data, model, or # of epochs, etc. 100 takes a while, make sure you have a good CUDA-enabled GPU (not like me). This one should train the model using the data you created earlier (the .txt) and then save the trained model in {your video id}.h5 in the appropraite folder.
At this point you're probably wondering how the videoId thing is pretty tedious. Yeah, don't think I didn't realize either. I plan to have a secure/personal front-end option for this, that doesn't use a bunch of CLI and videoId manual changes.