David and Matt discuss the challenges of constructing a dataset without test train and dev splits, highlighting the importance of considering dataset construction for better science and hyperparameter optimization. They reflect on the impact of oversampling and the trade-offs between dataset size and evaluation accuracy.
