⌘K
ENES
GitHub
conceptLearning Foundations~1 min readUpdated 2026-06-07#foundations#evaluation#data-leakage#validation

Train/validation/test splits & data leakage

You cannot judge generalization on data the model trained on. So you partition your data and keep some of it hidden — that held-out performance is your estimate of how the model behaves in the wild.

The three splits and their jobs

  • Training set — the model fits its parameters here.
  • Validation set — used to make decisions: hyperparameters, model selection, early stopping. You look at it many times.
  • Test set — touched once, at the very end, to report an honest number. If you tune against it, it stops being a fair estimate.

Cross-validation rotates the validation role across folds to get a more stable estimate when data is scarce.

Data leakage: the silent score-inflator

Leakage is any time information from outside the training set sneaks into the model, making validation scores look better than reality. It is the most common cause of "great in the notebook, terrible in production."

Common leaks:

  • Preprocessing before splitting — fitting a scaler/imputer/vectorizer on the whole dataset leaks test statistics into training. Fit on train only, then apply to validation/test (use a pipeline).
  • Target leakage — a feature that encodes the answer (e.g., a field only filled in after the outcome is known).
  • Temporal leakage — training on future data to predict the past. Time series must split by time, never randomly.
  • Group leakage — the same entity (user, patient, document) appears in both train and test, so the model "recognizes" rather than generalizes. Split by group.
  • Duplicate / near-duplicate rows straddling the split — rampant in scraped text corpora and a real concern for LLM benchmarks (contamination).

Rule of thumb

If a result looks too good, suspect leakage before you celebrate.

A realistic split that mirrors how the model will actually be used beats a clean random split that quietly leaks.

Connects to: overfitting · distribution shift · evaluation

Related notes

Learning FoundationsEvaluation metrics & what they hideAccuracy lies on imbalanced data. Precision, recall, F1, ROC-AUC — what each measures, and how to pick the metric that matches the cost of being wrong.Learning FoundationsMental Models for AI SystemsA compact set of mental models for thinking about AI systems beyond model demos.Learning FoundationsFeatures, representations & the curse of dimensionalityModels don't see the world — they see features. How representation quality decides everything, and why high dimensions are weird.Learning FoundationsGeneralization, overfitting & the bias–variance tradeoffA model is only useful if it works on data it never saw. Overfitting, underfitting, and the tradeoff that governs both.