How do you validate ML training data?

Direct answer

What a strong answer covers

How it works

Training data is validated against an approved baseline before any training job runs, because a model trained on broken data passes every code test. The checks cover schema and types, value ranges and allowed categories, missingness per column, uniqueness of the entity key, label availability and balance, temporal correctness so no feature is computed after its label timestamp, leakage heuristics such as features perfectly predicting the target, distribution shift against the previous training set, and coverage of protected or important groups. Violations block training at a severity chosen per check, and the validation report is attached to the run.

Example

The nightly training job for a churn model runs a validation step with TFDV-style statistics: it caught a week where the tenure column was in days instead of months after a source migration, a label rate that dropped from 4 percent to 0.3 percent when a downstream table stopped updating, and a new feature with 0.99 correlation to the label that turned out to be computed after churn. Each was blocked before a model could be trained on it.

Common mistakes

• Validating serving inputs but never the training set.
• Training on data with a label column that is partially unpopulated.
• Comparing against no baseline so a slow drift in the training data goes unnoticed.

A useful answer structure

Answer from your own work.

Do not memorise a script: use only examples you can explain honestly in follow-up questions.