Викторина, связанная с руководством · Средний Уровень

Викторина по передовому опыту разделения обучения, проверки и тестирования

Выбирайте разделы данных, которые отражают настройки развертывания, предотвращайте утечку между группами и временем и резервируйте тестовые данные для окончательной оценки.

Вопрос 1 из 8

How are validation data typically used during model development?

Предварительный просмотр каждого вопроса8 вопросы

Ответы и пояснения появляются по мере прохождения теста.

  1. How are validation data typically used during model development?
    • To fit the final model and report its unbiased score
    • To replace training data
    • To guarantee future performance
    • To compare choices such as hyperparameters without using the final test set
  2. What happens if a test set is repeatedly used to choose model variants?
    • It becomes more independent
    • It becomes part of model selection and the reported test estimate can become optimistic
    • It guarantees unbiased accuracy
    • It prevents overfitting
  3. When is a group-aware split appropriate?
    • Whenever data are already sorted alphabetically
    • Only when every label is unique
    • When related samples such as the same person or site must not appear across partitions
    • When the model has no validation set
  4. Why use a time-based split for a future forecasting task?
    • It trains on earlier observations and tests on later ones, matching temporal prediction
    • Это всегда увеличивает размер выборки
    • It randomly balances all dates
    • It removes seasonality
  5. Where should a scaler be fitted during cross-validation?
    • On all rows before splitting
    • Только на тестовом наборе
    • On a separate external dataset after evaluation
    • On the training fold, then applied to its held-out fold
  6. How should augmented versions of one original image be split?
    • Place every augmented version in the test set
    • Split them independently at random
    • Keep them in the same partition as their source image
    • Delete the original and split the copies
  7. Is a 70/15/15 split universally best?
    • Yes, for every data science task
    • No; partition sizes and method depend on data volume and evaluation goal
    • Yes, if a model is neural
    • No, because splits are never useful
  8. What does a held-out test score estimate?
    • Performance for the population and conditions represented by that test set
    • Performance for every future population with certainty
    • The probability that the model is fair
    • Скорость обучения