Split a dataset and check for leakage before training

Practice the sequence you already learned for “Split a dataset and check for leakage before training” until the order feels automatic. Drill ML-ops evaluation order from applied-ML training. The model is not trained until the split unit, time cutoff, and leakage checks are done.

1. Define the split unit (user, claim, time) so rows from one unit do not land in both sets.
2. Cut train / validation / test, using time if the problem is temporal.
3. Search for leakage: target in features, future timestamps, or ID hashes.
4. Fit only on the training set.
5. Report metrics on the held-out test set once.
6. Do not tune on the test set.
7. Do not shuffle a time series randomly.