Lesson · 12 min

Fix the Bias

Hands-on: rebalance a face-match training set before it ships.

Now reverse-engineer the problem. The lab runs a face-match model trained on a skewed dataset. Your job: rebalance the training data and watch subgroup accuracy converge — then decide whether the system is fair enough to ship.

How to play

Slide the skew control and watch the per-group error rates.
Find the fairness sweet spot — high accuracy for every group.
Read the fairness score and verdict.
Earn the Bias Hunter badge by fixing the dataset.

Big idea: fairness comes from the data you collect at the start, not from a switch an engineer flips at the end. XP: 100 on completion.

Hands-on lab

Bias Explorer Lab

A face-match model trained on a skewable dataset — watch fairness change live.

CONCERNING

Dataset skew70%

Mirrors the real worldOver-samples Group A

Training photos500

More data helps everyone — but balance matters more than size.

Fairness score

Accuracy gap between best & worst group:
13 percentage points

Concerning: one group is clearly getting worse results.

Group A395 photos · 50% of world

96% accuracy · 4% error ~20 wrong per 1,000 matches

data share (amber bar = how much of the dataset is this group)

Group B73 photos · 30% of world

91% accuracy · 9% error ~27 wrong per 1,000 matches

data share (amber bar = how much of the dataset is this group)

Group C33 photos · 20% of world

83% accuracy · 17% error ~34 wrong per 1,000 matches

data share (amber bar = how much of the dataset is this group)

What just happened?

The model's math never changed, only the data. Groups with fewer photos get more errors, and in the real world those errors become rejected loans, failed face unlock, and biased screening. Fairness gets decided when the data is collected.

Rebalance the dataset until every group's error is below 8% to claim the badge.

Knowledge check

The fastest lever to reduce group accuracy gaps is usually…

Tip: finish the lab first — completing it auto-completes this lesson.