Learn AI /How Machines Learn /Training example

Training example

Starter
How Machines Learn
Technical readers

Also called: sample, instance, training examples.

Detail

In five seconds

What it is

A single row of training data used to adjust a model, often made of input values and, in some methods, a target answer.

In plain words

A training example is one piece of data the system learns from, like one row in a spreadsheet or one labelled image.

An everyday example

A new bakery assistant learns the products by seeing one item at a time together with its correct name. One round pastry, shown with the label "bagel", is a single example. Its shape and appearance are the input, and the name is the target answer. The assistant needs many varied examples before reliably recognising new items. Likewise, one training example contributes to learning, but does not establish the whole pattern by itself.

Where the story stops working

Real machine-learning examples are usually turned into numeric representations before the model can use them, while the card story stays human-readable. Also, not every training example has a known correct answer: in unsupervised or self-supervised learning, the system may learn from structure in the inputs rather than from a back-of-card label.

Remember thisA training example is one observation presented to the learning process, often as input x paired with target y. It can influence the model alone or as part of a batch.

Understand it

Why it matters

If you can recognise what counts as one training example, you can ask better questions about quality, bias, coverage and privacy. For example: does each example represent a real case, is the label trustworthy, are important groups missing, and is one person or event counted many times? These questions affect whether a system learns useful patterns or misleading ones. The term also helps you separate three different things that are often mixed up: one example, the full dataset, and the trained model.

From one example to a trained model

Show that one training example is a single learning item, and many such items gradually change a model during training.

01
One item
Show a single example as one row: input text or values, and, if available, the target answer.
Draw one rectangular card split into two sections labelled Input and Target. Add a note below: 'One training example'.
02
Many items
Training usually uses many examples together, not just one.
Duplicate the card into a stacked set of rows. Label the stack 'Training data: many examples'.
03
Model update
Each example contributes a small update to the model's internal numbers during training.
Connect the stack to a central model box with looping arrows. Add small gear or slider icons and a label: 'Adjust internal parameters a little at a time'.
04
After training
The result is an updated model that can take new input and produce a prediction.
Show a second model box on the right labelled 'Trained model', with one new input card going in and one prediction card coming out.

A left-to-right flow diagram. It starts with one example shown as a single row, then zooms out to many rows, then shows repeated passes through a model, and ends with an updated model used on new input. Read from left to right to see how one item differs from the full collection and how repeated examples affect training.

Hover or tap each stage to highlight the difference between one example, the collection of examples, and the trained model produced after repeated updates.

Describe this without the diagram

The diagram explains that a training example is one individual learning item, not the whole dataset and not the model itself. First, it shows one row containing an input and, where relevant, a target answer. Next, it shows many such rows grouped together as training data. Then arrows lead to a model box, with loops indicating repeated updates from many examples. Finally, a trained model receives a new input and outputs a prediction. The message is that one example is a single case, many examples are used during training, and the output of training is an updated model.

With reduced motion: A static four-panel diagram shows: one example, many examples, repeated model updates, and a trained model making a prediction on new input. No animation is used; the loops are shown as fixed arrows.

A training example is one item a machine-learning system learns from. It might be one row in a table, one image, one short text, one sound clip, or one event in a log. A collection of these items is a dataset, but each individual item is a training example.

People also say sample or instance. In many introductions, these words are used almost interchangeably for one item in the data. But they can carry slightly different emphasis. Sample often points to something taken from a larger population, while instance often means one concrete case represented in the model's input format.

What goes in depends on the learning method. In supervised learning, one example usually contains input features, meaning the measurable parts the model receives, plus a label, meaning the target answer. In unsupervised learning, an example may have only the input features. During training, the system processes many examples and may update internal numbers so later it can produce predictions for new cases.

What comes out is a trained system. In many methods, this is a changed model rather than a stored answer for each example. A single example usually changes the model only a little. After many examples, patterns in the data can shape how the model behaves. This term is often confused with feature, which is one column or property inside an example, and with dataset, which is the full collection of examples.

How “Learning the pastry counter” maps onto the real thing

In the story	In the AI system
one index card with a query and the correct document	one training example with input and, in supervised learning, a label
the words on the front of the card	the input features presented to the model
the note on the back naming the relevant document	the label or target answer
the whole box of cards	the dataset

A worked example

For the audit file, the desk officer shows how one reviewed request contributes to training a classifier that routes messages as either programme or administration.

Starting from: One example: subject line = "Comments requested on draft workplan"; reviewed target = "programme"

Convert the permitted words or other attributes into an input vector, conventionally written as x.
Pair that vector with the reviewed target y, in this case "programme".
Have the current model calculate its predicted category or category probabilities.
Compare the prediction with the target and calculate the error for this example.
Use that error, alone or combined with errors from a batch, to adjust the model's parameters slightly.

Result: The example does not create a fixed rule that every message mentioning a workplan is programme-related. It contributes a small adjustment alongside many other examples, from which the model learns broader statistical patterns.

In a United Nations workplace

A desk officer in a Resident Coordinator's Office is closing an internal audit query about a classifier that routes incoming requests to the correct focal point. The file needs a concrete account of the training data. One training example is one previously reviewed request, represented by permitted text features, paired with the routing category assigned by staff.

The example may be processed alone or in a batch. Its features and reviewed target help determine a small parameter adjustment. The audit response can therefore distinguish one training example from the full dataset, which is a useful distinction when the annex is due by close of business.

Before you try thisRequests may contain names, contact details or confidential material. They must not be copied into a tool that has not been approved for that information. Use only necessary fields, follow the office's data protection guidance, restrict access and retain human review of routing decisions.

Where you may hear it

An internal audit exit meeting reviewing model documentation
A vendor demo to an ICT governance board
A data-lab session preparing labelled records for quality review
An evaluation office briefing on a predictive model's methodology

Avoid confusion

Often confused with

Common misconceptions

Common misconception: A training example is the same as the whole dataset.
Correction: A dataset is the collection. A training example is one item inside it, such as one row, one image or one text. Confusing the two makes it harder to judge how much data a model actually learned from and whether the data covers enough cases.
Common misconception: A training example always includes the correct answer.
Correction: That is true in many supervised-learning setups, where an example has inputs and a label. But in unsupervised, self-supervised and some other methods, examples may not have a human-provided target answer. The system may instead learn patterns, structure or useful representations from the inputs themselves.
Common misconception: One example teaches the model a complete rule.
Correction: Usually one example only causes a small update. Models typically need many examples because they are estimating patterns across data, not copying a single case into a reliable general rule. If one example has too much influence, that can be a sign of overfitting or data imbalance.

Quick check

Which statement best describes a training example?

Go deeper

The simple version

A training example is one piece of data the system learns from, like one row in a spreadsheet or one labelled image.

The technical version

Technically, a training example is one observation used by a learning algorithm. In optimisation-based methods, it is often represented as an input vector x and, in supervised learning, an associated target y. The terms sample and instance are often used for the same unit, though sample may emphasise that the observation was drawn from a broader population, and instance may emphasise one concrete case in feature space. Examples may be processed singly or in batches, and each can contribute to an update of model parameters. The usual result is an adjusted model, not a permanent one-to-one lookup table of the examples.

Reviewed external resources

EncyclopaediaWikipedia search: Training example (opens in a new tab)Wikipedia
This search results page helps readers compare different Wikipedia entries and related terms around training examples, and it suits people who want a broad starting point rather than one curated article.
search results, not one pagelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course gives a structured introduction to core machine-learning ideas and suits beginners who want training examples explained in the context of how models are built.
Intermediatelink checked 2026-07-31
VideoStatQuest (opens in a new tab)StatQuest with Josh Starmer
This video source is useful for learners who prefer step-by-step spoken explanations and visual intuition before reading more technical material.
Starterlink checked 2026-07-31
VideoVideo explanations: Training example (opens in a new tab)YouTube search
This search results page helps readers find multiple video explanations of the term and suits people who want to compare teaching styles rather than rely on one source.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Training example (opens in a new tab)Reddit search
This search results page shows community discussions and practical questions about the term, and it suits readers who want to see how people informally use and debate the wording.
search results, not one pagelink checked 2026-07-31

Where this came from

Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

The one thing to remember

A training example is one observation presented to the learning process, often as input x paired with target y. It can influence the model alone or as part of a batch.