What is model evaluation?
Model evaluation is how you find out whether a model actually works. It combines quantitative metrics, computed on data the model did not see during training, with qualitative inspection of individual predictions to build a full picture of performance. A single headline metric rarely tells the whole story, so good evaluation looks at metrics per class, per slice of data, and on the specific cases that matter.
Done well, it answers not just how good a model is, but where and why it fails, which is what tells you how to improve it.
Key takeaways
- Evaluation uses held-out data and multiple metrics, not one number.
- It pairs quantitative metrics with qualitative error inspection.
- Slicing results by class and data segment reveals hidden weaknesses.
How it works
A model is scored on a validation or test set using task-appropriate metrics, then results are broken down by class and by data slice to expose uneven performance. Comparing predictions against ground truth surfaces systematic errors, mislabeled data, and edge cases. This diagnosis, rather than the aggregate score alone, guides the next round of data or model changes.
Why it matters
Most of the gains in a mature machine learning project come from understanding failures, not from chasing a higher aggregate score. Rigorous evaluation catches problems like data leakage, label errors, and blind spots on rare cases before they reach production, and it is the foundation of a data-centric approach where improving the data is often the fastest path to a better model.
Frequently asked questions
What is the difference between a metric and evaluation?
A metric is a single measurement, such as precision. Evaluation is the broader process of choosing metrics, slicing results, and analyzing errors to understand a model.
Why isn't accuracy enough?
Accuracy hides class imbalance, per-slice weaknesses, and the specific failure cases that matter most, which is why evaluation looks deeper.
Related terms