Foundations

Training interview questions

Training questions check whether you understand why a model that looks good can still be wrong. The classic opener is a headline number, like 99% accuracy or a near-perfect validation score, and the question is whether to believe it.

Interviewers want to hear how you’d set up validation so it reflects how the model will really be used, what regularisation does to the fit, how you’d tell overfitting from a leaky split, and what to do when the class you care about is 1% of the data. Naming the bias-variance trade-off isn’t enough: expect to be asked what you’d change and what you’d expect to see.

Try one now

Friendly

A model hits 99% accuracy on fraud detection. Is that good?

Free, no sign-up. The interviewer asks follow-ups on your reasoning, then rates your answer.

What the 20 questions cover

  • Cross-validation9
  • Class imbalance (resampling, class weights, thresholds)8
  • Bias-variance trade-off4
  • Overfitting and underfitting4
  • Regularisation (L1, L2, elastic net)4
  • Gradient descent2
  • Optimizers (Adam, momentum, RMSprop)2
  • Hyperparameter tuning (grid, random, Bayesian search)2
  • Early stopping1

A question can cover more than one topic, so the counts add up to more than 20.

How the practice works

  1. The AI interviewer asks the question and follows up on your answer, the way an interviewer would.
  2. Stuck? Ask for a hint. Each one gives a little more away, so you can get unstuck without being handed the answer.
  3. Get a written debrief. Four ratings, each from Needs work to Interview-ready, tied to what you actually said, and what to study next.

More foundations questions

See all categories