This is my attempt at the toxic competition on Kaggle by Jigsaw.
The notebooks are separated by their type of pretrained embedding models. Preprocessing can be found in helpers.py and models can be found in architecture.py.
I've taken my best results from every type of model and collected them into an ensemble, then applied the Toxic Avenger technique by the1owl. I do not recommend putting too much thought into the weights I've chosen, since they're based off of the success of my model with the Public Leaderboard, which could have a different distribution than the Private Leaderboard or any other dataset.