Skip to main content
Tokenline

Line 1 · Station 3 of 6

Training Data: You Are What You Learn From

Discover why the amount, quality and variety of training data decide how well — and how fairly — a model works.

  • 6 min
  • +50 XP
  • 2 knowledge checks

By the end of this station, you can:

  • Explain why data quality matters as much as data quantity
  • Identify when a dataset doesn't represent the people or cases it will be used on
  • Connect gaps in training data to unfair or unreliable results

A model only knows its examples

A machine learning model learns everything it knows from its training data. It has no other way to learn about the world.

So if the data is messy, missing important cases, or full of mistakes, the model learns those problems too. People sum this up as “garbage in, garbage out.”

Three things that make training data good

  • Quality — examples are accurate and labels are correct.
  • Quantity — there are enough examples for the model to find real patterns, not flukes.
  • Representation — the data covers the full range of people, places and situations the model will face.

Key terms

Training data
The collection of examples a machine learning model learns from.
Dataset
An organized collection of data, such as a table of songs with their features and labels.
Representative data
Data that reflects the full variety of cases a model will actually be used on.
Garbage in, garbage out
The idea that a model trained on flawed data will produce flawed results.
ExampleThe sunny-day problem

You

A team trains a model to spot stop signs using thousands of photos — all taken on clear, sunny afternoons.

AI assistant (sample)

The model scores great in testing. Then it struggles at night, in rain and in snow.

Why it matters: The model wasn't “bad at stop signs.” It never saw stop signs in those conditions. Its training data didn't represent the real world it was used in.

Knowledge check 1

+10 XP on first try
A speech-recognition app was trained mostly on recordings of adults. Kids say it often mishears them. What's the most likely cause?

Interactive lab

Data Diet

You're training “Leaf Doctor,” an app that spots sick plants from photos. Choose which datasets go into its training data and watch how well it does on each kind of test photo.

Training datasets

Test results (simulated)

  • Bright daylight photos88%
  • Low-light photos58%
  • Blurry phone photos56%
  • Unfamiliar plant varieties55%
Overall accuracy64%

1,200 training photos selected

Weak spot: unfamiliar plant varieties (55%). The model has barely seen photos like these.

What to notice: More photos isn't automatically better. Duplicates add nothing new, wrong labels make things worse, and any situation missing from the training data becomes a blind spot.

Learn moreWhere does training data come from?

It depends on the model. Some datasets are collected on purpose, like labeled medical images. Others are gathered from large amounts of public text and images online.

Labels are often added by people, which takes time and can introduce mistakes or personal judgment. That's one reason datasets need to be checked carefully.

Knowledge check 2

+10 XP on first try
You're building a model to recommend clubs to all students at your school. Which dataset would likely work best?

Key takeaways

  • Models learn only from their training data — its strengths and its flaws.
  • Good data is accurate, plentiful and representative.
  • Missing groups or situations lead to unreliable results for those cases.
  • Biased data can produce biased AI.

Sources for this lesson

  • Google for Developers. Machine Learning Crash Course. Full citation
  • National Institute of Standards and Technology (NIST). Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, 2023. Full citation

Finish this station

Answer every knowledge check (right or wrong) to unlock completion.

  • Knowledge check 1 — not answered yet
  • Knowledge check 2 — not answered yet

0/2 checks answered