How to Win a Data Hackathon (Hacklytics 2021)

Eugene Yan · 5 years ago

A mentor at Georgia Tech's 36-hour Hacklytics 2021 datathon identified patterns among winning teams, including their use of publicly available datasets and APIs rather than web scraping, reliance on pre-trained models and libraries like BERT and VADER for machine learning tasks, and deployment of working prototypes with user interfaces. The winning teams used resources such as Kaggle datasets with 40,000+ records and pre-trained models fine-tuned on hackathon data rather than training models from scratch. Success depended on rapid prototyping, UI development with React or Streamlit, and deployed demos that judges could interact with rather than novel machine learning research.