Day 5 · Good Data vs Bad Data

🏷️ The pro wordLopsided training data creates bias. Too few examples of one class is class imbalance. An AI that only works on data like its training set has poor generalization.

🍎 Lab · The Fruit Sorter

Teach a robot to sort fruit

The robot sorts 🍎 apples from 🍋 lemons using two clues: size (left→right) and yellowness (bottom→top). The colored zones show what the robot would guess anywhere. The little dots with rings are the test — fruit it's never seen. An ✗ means it got that one wrong!

⬆ more yellow

bigger size ➡

🤖 Test accuracy

—

Drop your own examples — pick a fruit, then click the box:

Or load a situation:

Try it: tap 🕶️ Biased, then ❌ Mislabeled, then ✅ Good data. Watch the accuracy number jump around — same robot, different data!

🕶️ Why It Matters · Bias

The sneaky problem: bias

Load 🕶️ Biased above. See how the robot only ever saw apples from one tiny corner? It looks sure of itself — but any apple that's a bit different gets called a lemon (all those ✗s). This is bias, and it's the #1 way real AIs go wrong.

⚖️ In real life this matters a LOT: a face AI trained on only one kind of face, or a voice AI trained on only one accent, fails everyone it never saw. Fair AI needs fair, varied data. That's a rule real engineers live by. 🛡️

🔧 Boss Build · Fix the AI

You're the data doctor

Tap 🕶️ Biased to break the robot, then go back to the lab and add your own good examples (spread them out to cover where real apples and lemons actually live) until you push the accuracy to 90%+. Fix it with data!

🎯 Mission: get test accuracy to 90% or higher. Current best: 0%

🚀 Beyond the Basics

How the pros really do it

Eyeballing your data is step one. Professionals treat data as the main job (data-centric AI):

🟢

Spot bad data by eye (what you did)

✅ Catches obvious problems. ⚠️ Misses subtle, hidden bias.

⚖️

Balance and fairness tools

Re-sampling / re-weighting classes, and fairness audits across groups. ✅ Fairer models.

📋

Document it (AI practice)

Datasheets and model cards record the data used and known limits - required at real companies.

🏷️ Pro names to look upbias · class imbalance · data-centric AI · fairness audit · model cards.

🎉 Day 5 Complete!

You learned the golden rule of AI: garbage in, garbage out. Good, fair, varied data makes a smart AI — bad data breaks it, no matter how fancy it is.

Day 6: Scoring an AI →

Tomorrow: how do we actually measure if an AI is any good? You'll meet true & false positives, precision, recall, and the famous F1 score — as a game. 🎯