Glossary term
Glossary term
Training and Fine-Tuning
Data used during development to tune, compare, or select models. It should be separated from training data where appropriate to reduce misleading performance claims. Validation data should be protected from leakage into training and should represent relevant subgroups, edge cases, and operating conditions.
The MLPerf benchmark suite enforces strict separation between training, validation, and test sets to ensure fair performance comparisons.
Kaggle competitions enforce hidden validation and test splits to prevent submission overfitting and to support reproducible evaluation.
ImageNet's standard 50,000-image validation split has been used since 2010 to tune and compare image classification models.