Data Is the New Recipe

Garbage in, garbage out — the oldest law of machine learning.

A model trained on bad data makes bad predictions with great confidence. That is why data work is most of the job in real ML projects. Think of data as the recipe and the model as the cake: no recipe tweak fixes rotten eggs.

What makes data "good"?

Accurate — labels actually match reality.
Representative — it looks like the world the model will meet.
Balanced — enough examples of every category you care about.
Clean — duplicates, typos, and mislabeled rows removed.

The next lessons make this personal: you’ll train a real classifier and then watch what happens when the data behind it is skewed.

Try it yourself

Try it — spot

Myth or truth? Data-quality edition.

"More data always means a better model."

"A model can only learn patterns its data contains."

"Duplicates make a model fairer by reinforcing important examples."

"Data quality is most of the work in real ML projects."

Knowledge check

A dog-vs-cat classifier is trained on 990 dog photos and 10 cat photos. What will happen?

Your training photos are all taken in bright sunlight. Users complain the model fails at night. What’s the problem?