Data & Datasets
A model is only as good as its data — and only as trustworthy as its measurement. The universal ritual: split your labelled examples into a TRAINING set the model learns from and a TEST set it never sees, the held-back exam. The test score is your only honest preview of real-world performance.
Slide the split: more training data improves the model, but leaves fewer test examples — shrink the test set too far and the score itself turns into noise. Then flip the leakage toggle: let test examples sneak into training, and accuracy leaps to a glorious, meaningless 99%. Leakage is the most common silent bug in ML — and the same disease as benchmark contamination in the evals lesson.
reported accuracy: 87%
The test set is the held-back exam: the model never sees it during training, so the score predicts real-world performance.
Check yourself
Your model scores 99% in the lab and 70% in production. What is the first thing you should suspect?
Go deeper (free): Google ML Crash Course — Working with data ↗