Skip to content
BENIBIX Learn

Model Evaluation

AI📏 Evaluation & BenchmarksEvaluation Basics

Also known as ML Model Evaluation, Model Testing

Checking how good a model is. You test it on data it has not seen.

Think of it like

It is like a test at school. The teacher checks what you learned with questions you have not seen.

Example

Jack trains a model to spot sick plants from leaf photos. He tests it on 1,000 new photos, and it gets 912 right, so its accuracy is 91.2%.

Why it matters

A model can look great on the data it learned from and still fail on new data. Evaluation shows which one you have.

Definition

Model evaluation is how you measure how well a trained model does its job. You run it on data it never saw while it learned, and compare its answers with the right ones. The result is a score, such as how often it is right. The kind of score depends on the task, so a spam filter and a chatbot are judged in different ways. Without this step, you cannot tell a good model from a lucky one.
The model answers new questions, each answer is checked, and the checks add up to a score.

Common mistake

Testing on the same data the model trained on. It has seen those answers, so the score will be far too high.

Key takeaways

  1. 1
    New data. Never seen in training.
  2. 2
    Check against truth. Right or wrong.
  3. 3
    Fit the task. Pick the right score.