News classifier, fine-tuned BERT
700K+ articles, 30 categories, tuned to 98%.
A news-article classifier across 30 categories on a 700K+ article corpus. I preprocessed and normalized the raw dataset, then parallelized tokenization with mpi4py to multi-thread the ingest on my local machine before training. Fine-tuned a distilled English BERT for three epochs on Kaggle's GPU, and pushed accuracy from 70% to 92% to a final 98% through hyperparameter tuning. That score placed in the top 10 on the held-out test leaderboard.