Data leakage occurs when information unavailable at prediction time influences training or evaluation. It can make a model appear accurate in a notebook while failing on genuinely new examples.
Review the split
Ask whether rows from the same person, event, or time period appear in both training and test sets. Grouped or time-based data may need a specialized split rather than a random row split.
Fit transformations on training data
Imputation, scaling, feature selection, and encoding should learn parameters from training data only. A pipeline helps apply the same fitted transformations consistently without using the held-out set to shape the model.
Use a task-relevant metric
Document the metric definition, baseline, class balance or error scale, and sample size. Consider whether false positives and false negatives have different consequences.
Report what was not tested
State the dataset scope, collection period, known missing groups, and limits on generalization. A single test score does not establish fairness, causation, or suitability for high-impact decisions.