The model's entire world
A model never sees the real world. It only sees its training data, so that data is its whole world. Training data is a collection of examples, and a collection like that is called a dataset. For a spam filter, the dataset is thousands of emails, each marked spam or not spam. For a music app, it is what millions of people played and skipped. For a chatbot, it is a giant pile of text from the internet and books. Whatever is common in the data becomes normal to the model. Whatever is missing, the model simply does not know exists.
More is good, better is better
You might think the only rule is "more data". More does help, and modern models eat enormous amounts. But quality matters more than quantity. Imagine a dataset where one photo in ten is marked wrong: a dog marked cat. The model will learn from those mistakes just as faithfully as from the correct ones. Variety matters too. Ten thousand clean, varied examples beat a million messy copies of the same thing. Programmers have a blunt phrase for this: garbage in, garbage out.
A score, then a line
Many models are classifiers: they sort things into groups, such as spam or not spam. A classifier does not answer with a plain yes or no. It gives a score, something like "this email looks 83 percent spam-like". Someone has to turn that score into a decision. The tool for that is a threshold: a line, drawn at a chosen score. Anything above the line is called spam and goes to the junk folder. Anything below goes to your inbox. Move the line and the decisions change, even though the model did not.
Two ways to be wrong
| Mistake | What it means | In your inbox |
|---|---|---|
| False alarm | Flagging something harmless | A friend's party invite lands in spam |
| Miss | Letting a bad one through | A scam email reaches you |
| The trade-off | Fewer of one means more of the other | Where you draw the line decides which you get |
A smoke alarm
A smoke alarm is a classifier with a threshold. Set it very sensitive and it screams at burnt toast: lots of false alarms. Set it lazy and it stays quiet through a real fire: a miss. No setting gives you zero of both. You choose which mistake hurts less. For a smoke alarm, false alarms are annoying but a miss is deadly, so you keep it sensitive. A spam filter makes the opposite choice, because losing a real email hurts more than seeing one junk message.
Bad data teaches bad habits
If the examples are lopsided, the model copies the lopsidedness. A face-unlock model trained mostly on one group of people works worse for everyone else. It is not being unfair on purpose. It simply never learned the rest. This is called bias, and lesson 7 returns to it. Next: where inside the model does all this learning actually live?
Check yourself · 3 quick questions
Why is training data described as the model's whole world?
Show answer
Because the model only ever sees those examples, and nothing else — A model never sees the real world; whatever is common in its data becomes normal to it, and whatever is missing it does not know exists.
A spam filter moves its threshold so that fewer real emails land in spam. What else changes?
Show answer
More spam gets through to the inbox — Fewer false alarms means more misses: moving the line trades one kind of mistake for the other.
A face-unlock model works well for one group of people and badly for others. What is the most likely cause?
Show answer
The training data was lopsided toward one group — A model copies whatever its data looks like; if one group was rare in the examples, the model never properly learned them. That is bias.