Moj Multilingual Abusive Comment Identification - Challenge

Submission for the Moj Abusive Comment Detection Challenge, hosted on Kaggle.

Setup

1. Download the dataset and extract it the root directory.
2. Create a virtual environment:

python -m venv env # virtual env
pip install -r requirements.txt
source env/bin/activate

3. Change index in main.py to choose which model of the ensemble to train/test.
4. Run python main.py
5. utils.py contains helper functions for caching and ensembling.

Note: BERT models are large, GPU with 16GB VRAM required. Batch size can be reduced if training on 8GB GPU.