⚡ Activation

💭 in your head · nothing to type

You fine-tuned a model and asked it about three examples from your dataset. All three answers are perfect, word for word as in the labels.

The first reaction is joy. The second, if you listen closely, is suspicion.

What exactly did this check show: that the model learned the task — or that it memorized the answers?

👁 Reflection

💭 in your head · nothing to type

You know the difference between "understood" and "memorized" from school. A student who learned the answers to the exam cards will ace those cards and fail on a new question.

It's the same with a model. Checking on training examples measures memory, not skill.

Consider: which examples from your task could the model definitely not have learned from?

💡 Concept

Holdout: set it aside before you start

The main rule of evaluation: some examples are set aside before training and never shown to the model. This is the holdout set.

Data is usually split into three parts:

• train — the model learns from these;
• validation — you tune settings on these along the way;
• test — you look at it once, at the end.

If you edited the data while looking at errors on test, it has become validation. An honest evaluation needs a new one.

Overfitting