How many labeled examples do I need?

Enough to cover the territory with the rare cases represented; simple problems run on thousands, complex ones on far more, and coverage beats raw count.

What is the difference from an unlabeled example?

The unlabeled case has inputs but no known answer; it can be scored by a trained model but cannot teach one directly.

Can models train on partly wrong labels?

They absorb the error rate they are fed, which is why auditing the collection is worth more than another algorithm tweak.

AI Glossary / AI and machine learning / Beginner

Labeled example

Definition

A labeled example is a single training case with its known answer attached: the inputs plus the label, the atomic unit of supervised learning.

What it means

A model learns nothing from raw data alone. It learns from cases where the outcome is known: this email was spam, this customer did churn, this image did contain a crack. Each of those answered cases is a labeled example.

Training is the digestion of many labeled examples: the model adjusts itself to reduce disagreement with the answers, and its eventual skill is bounded by how many it saw, how well they covered the territory and how often the answers were right.

The word is worth separating from its neighbors: the example is the case, the label is the answer, and the labeled example is the pair, ready for the training set. An unlabeled example is the same case awaiting, or lacking, its answer.

Counting labeled examples is the first sizing question of any project: dozens teach almost nothing, thousands teach something, and the coverage of the collection, which kinds of cases got answered, matters as much as the raw count.

For a business, labeled examples are the inventory the model is made of. Their cost, quality and coverage deserve the same management attention as any other raw material, because the finished model can contain no more truth than they carried in.

How it works

Define the unit: what one example is, a customer, a transaction, an image, and what its label means and when it becomes known.

Collect labeled examples across the territory the model will serve: all classes, all segments, all seasons that matter.

Check the balance: rare outcomes need deliberate oversampling in the collection or they vanish from the model's education.

Version the collection: which examples, labeled how, when, so any model can be traced to the exact inventory it learned from.

Audit samples of the collection for label accuracy, and feed corrections back into the pool.

Three practical examples

A fraud model trained on 40,000 labeled examples, of which 900 are confirmed fraud, with the rarity handled deliberately.
A medical project whose labeled examples cover three hospitals, so the model is not educated in one hospital's habits alone.
A versioned training collection where every model release names the exact set of labeled examples it digested.

Think of it this way

Think of a phrasebook for a language learner: each entry pairs a situation with the right thing to say. One entry teaches one situation; a well-organized book of thousands, covering restaurants, stations and arguments, teaches the language. Labeled examples are the entries.

The limit: the book covers what it covers. A learner meeting a situation the book never covered, a region, a class, an era, improvises, and the model's improvisation in uncovered territory is exactly as unreliable.

Worked example

counting what the model ate

Suppose a retailer reviews its demand model's training inventory. The figures are illustrative.

Of 200,000 labeled examples, 70 per cent come from two flagship regions and one busy quarter, leaving whole territories thin.

A targeted collection effort, pulling and labeling history from the under-covered regions, lifts national accuracy by six points without any algorithm change.

The finding reframes the roadmap: the next quarter's budget goes to inventory coverage, not to a fancier model.

Illustrative case study

A fictional bank trained on a labeled example collection inherited from a predecessor system, and nobody noticed the collection contained almost no declined applications, because the old system had only logged approvals.

The model learned a world where nearly everyone is approved, and its early live behavior reflected that education until the gap was found and the collection rebuilt.

The invented lesson: the model is what it ate. Audit the labeled example inventory for coverage before auditing the algorithm for cleverness.

Common misconceptions and mistakes

Any answered case is a good labeled example. Coverage, correctness and timing all matter; an answered case from the wrong era or a biased process can teach the wrong lesson.
More examples always help. More of the same examples help little; marginal value lives in uncovered territory and corrected errors.
The label is all that matters. The inputs must also be right: a correct answer attached to corrupted inputs teaches a corrupted lesson.

Business use

Maintain the labeled example collection as a managed asset: counts by class, segment, era and source, with an owner.

Spend the next labeling budget where coverage is thinnest, not where collection is easiest.

Keep the collection versioned so every model can be traced to exactly what it learned from.

Check your understanding

Why does the coverage of the labeled example collection matter as much as its size?

Answer: Because the model learns only the territory the collection covers, so thin regions, classes or eras become unreliable predictions regardless of how large the total count is.

Related terms

Learn next

Be Productive Using AI covers asking what an AI system actually learned from before trusting its answers: https://aiforbusinessandlife.com/academy/be-productive-using-ai. [C1]

Sources

[S2] Google, Machine Learning Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Labeled data point Next: Labeled precision