Data Types for AI Models

Playback speed

Swipe, scroll or use ← →

Labels: does the example come with its answer?

Before choosing a method or a service, describe the data you actually hold. The first question is the cheapest one to answer and the most consequential: does each example arrive with the answer attached?

Labeled data pairs each training example with an explicit label or target — an image marked cat, or a row of sales data with a target value. Supervised learning uses those known targets to learn a mapping from the input attributes to the answer you want to predict.

Unlabeled data contains input features without labels or target values. A model working with unlabeled inputs has to discover patterns, such as groupings, from the inputs themselves, because there is no supplied answer to learn against.

1 / 20

Labels: does the example come with its answer?

Labeling is a property of the examples, not of their file format. The same content can exist in either state.

Content	Labeled version	Unlabeled version
Product photographs	Each photo carries a category caption or annotation	Photos supplied with no annotation at all
Sales table	A target column holds the value to predict	Only input columns, no outcome
Customer emails	Each message tagged with its resolved intent	Raw messages, nothing tagged

So a table can be labeled when it includes a target column and unlabeled when it holds only input columns. Adding a timestamp, changing the file format, or switching from text to images changes none of this: only the presence of a target does.

2 / 20

Quick check

What makes a training example labeled?
1. AIt carries an explicit label or target that supplies the known answer
  
  Right. A label or target is the answer attached to the example, and supervised learning maps the inputs to it.
2. BIt carries a timestamp, although the value to be predicted is left out
  
  A timestamp records when something was measured; without a target value the example still has no answer.
3. CIt groups similar inputs together without relying on any predefined outcome
  
  Discovering groups from similarity is what happens when the answers are missing, so it describes unlabeled data.
3 / 20
Structure: does the data follow a fixed schema?

The second question is about organization, and it is independent of the first.

Structured data is organized in tables or fixed schemas, where relationships are represented through fields, rows, and columns. A warehouse table, a CSV export, and a database view are all structured.

Tabular data is that structured arrangement in its most familiar form: each row can represent a sample and each column can represent a feature or a label.

Unstructured data is content such as free-form text, images, audio, and video, which is not organized as a fixed table schema.

4 / 20
Structure: does the data follow a fixed schema?

Rows and columns carry a consistent meaning that is worth learning once:

Element What it usually holds

Row One sample: a single observed entity

Feature column One measurable property of that sample

Target column The label the model should predict

A useful consequence: a fixed CSV schema stays structured and tabular even when one of its fields contains free-form text, such as a support message. The container is a table; only that one field's content is unstructured.

5 / 20
Quick check

Which description matches structured data?
1. AVisual content that carries captions but cannot be placed inside a dataset
  
  Images are unstructured content, and captions are labels; neither prevents data from being stored or used.
2. BObservations that carry no labels at all and therefore cannot be arranged into columns
  
  Label state and structure are separate properties: an unlabeled dataset can still have rows and columns.
3. CData organized in tables or fixed schemas, with defined fields, rows, and columns
  
  Right. Structure is about organization into fields, rows, and columns under a fixed schema.
6 / 20
Modality: what kind of content is it?

Modality names the form of the content itself, which is a third and separate axis.

Text data includes language content such as web pages, books, articles, documents, and emails.

Image data represents visual content, used by classifiers and by generative vision models. Images may also be paired with captions or labels, and those captions are what make an image dataset labeled.

Modern generative systems can work across several modalities at once — text, code, images, audio, video — while a traditional model usually specializes in one type at a time: an image classifier cannot write text, and a sentiment model cannot draw.

Text and images are modalities, while structured and unstructured describe organization, so a modality never tells you whether the data is labeled.

7 / 20
Quick check

Which pairing identifies two modalities correctly?
1. ABooks and emails are timestamps, while photographs and diagrams are target columns
  
  Timestamps record measurement time and target columns hold answers; neither describes the content itself.
2. BBooks and emails are text data, while photographs and diagrams are image data
  
  Right. Language content such as books, articles, and emails is text, and visual content is image data.
3. CBooks and emails are tabular targets, while photographs and diagrams are fixed schemas
  
  A target and a schema are properties of a dataset's organization and labeling, not the kind of content it holds.
8 / 20
Keep your progress in the app

That’s 3 of 8 quick checks. In the app they stay answered, and every lesson remembers where you left off.

Get the app Keep reading here
Time: are the observations ordered in time?

Time-series data is a collection of observations or measurements recorded over regular intervals of time. Each observation is associated with a specific timestamp or time period, creating a sequence of data points ordered chronologically.

That ordering is what makes the data type distinctive: shuffle the rows of a customer table and you lose nothing, but shuffle a time series and you destroy the pattern the model exists to learn.

9 / 20

Element	What it usually holds
Row	One sample: a single observed entity
Feature column	One measurable property of that sample
Target column	The label the model should predict

Time: are the observations ordered in time?

A forecasting table needs three columns at minimum:

Column	Role
Item identifier	Which item this measurement belongs to
Point-in-time target	The value of that item at that moment, and what the model predicts
Timestamp	The date and time the value was measured

Additional columns can carry other factors that may influence the forecast — units sold, store location, inventory levels, or covariates such as weather.

Note what this implies: a time series is usually tabular too. Chronological and tabular are compatible descriptions, not competing ones.

10 / 20

Quick check

Which feature is essential to the definition of time-series data?
1. AEach observation has a timestamp or time period and is ordered chronologically
  
  Right. Timestamped observations arranged in chronological order are what define the type; modality and label state can still vary.
2. BInputs are grouped by similarity, and the time of measurement is deliberately discarded
  
  Grouping by similarity describes what a model does with unlabeled inputs, and discarding time would remove the defining property.
3. CEvery observation is an image with a hand-written caption and no timestamp or time period
  
  Modality and labeling are separate axes, and an example without a timestamp cannot be part of a time series.
11 / 20

Four questions, four independent answers

The four properties are separate axes, so one dataset carries several descriptions at once. Ask them in order and the description writes itself.

Ask	Answers
Label state — do known answers accompany the examples?	Labeled / unlabeled
Structure — do the fields follow a fixed schema?	Structured / unstructured
Modality — what form is the content?	Text, image, audio, video, …
Time — are timestamps and order essential?	Time-series or not

Exam scenarios exploit the confusion between axes, so watch for claims that mix them: that timestamps prevent a row from having a target, that text content makes a CSV unstructured, or that a caption turns a photograph into text.

12 / 20

Quick check

Which comparison of the properties is accurate?
1. ALabeling describes rows and columns, while structure describes whether the targets are correct
  
  Rows and columns belong to structure, and label correctness is a data-quality question rather than a definition.
2. BLabeling would describe chronological order, while structure would describe whether an image has a caption
  
  Chronological order is the time axis, and a caption is a label on an image, so both halves name the wrong axis.
3. CLabeling says whether targets are supplied; structure says whether the data follows a fixed schema
  
  Right. The two axes answer different questions and can be combined freely with modality and time.
13 / 20
Worked example: a forecasting table

A dataset stores one row per product and per day. Its columns are product ID, timestamp, the sales value to predict, and the weather that day.

Run the four questions:
- Structure — fixed columns, one row per observation, so it is structured and tabular.
- Time — every row carries a timestamp and the rows are chronological, so it is time-series data.
- Label state — the sales value is the target to predict, so the examples are labeled.
- Modality — the fields are identifiers, numbers, and dates rather than free-form media.
All four descriptions hold at once. Weather is an extra factor that may influence the forecast — a feature column, not a change of modality.
14 / 20
Quick check

That forecasting table has product ID, timestamp, a sales value to predict, and weather. How is it described?
1. AIt is image data and labeled, because weather is a visual feature in every row
  
  Weather arrives as a data field here, not as a picture, so nothing in the table is image content.
2. BIt is unstructured and unlabeled, because timestamps stop rows from holding targets
  
  Timestamps and targets sit on different axes; a row can carry both, and fixed columns keep the table structured.
3. CIt is tabular, structured, time-series, and labeled, because the sales value is its target
  
  Right. Rows and columns give structure, timestamps give the chronological order, and the sales target makes it labeled.
15 / 20
Worked example: photos, emails, and a CSV

A team holds product photographs paired with category captions, and separately a collection of free-form customer emails with no targets.
- The photographs are image data, and their captions are labels, so they are labeled.
- The emails are text data. They are free-form rather than fixed-schema, so they are unstructured, and with no target they are unlabeled.
16 / 20

Worked example: photos, emails, and a CSV

Now a harder container. A CSV holds one row per support ticket, with fixed columns for account and region, a free-form message field, and no outcome column.

Question	Answer	Why
Structure	Structured and tabular	The CSV has a fixed schema of rows and columns
Modality	The message field holds text	Free-form language inside one field
Label state	Unlabeled	No outcome or target column exists

One free-form field does not dissolve the schema around it, and language inside a column does not turn the file into unstructured media. Judge the container and the examples separately.

17 / 20

Quick check

A CSV has fixed account and region columns, a free-form message field, and no outcome column. What is the most precise description?
1. AThe CSV is tabular and structured, the message field holds text, and the examples are unlabeled
  
  Right. A fixed schema keeps the file tabular and structured, while the missing outcome column leaves the examples unlabeled.
2. BThe CSV is time-series only, the message field is the target, and the examples are labeled
  
  No timestamp column is mentioned, and a free-form message is an input field rather than the value to predict.
3. CThe CSV is unstructured image data, the message field is a timestamp, and the examples are labeled
  
  A CSV is not image content, a message is not a timestamp, and nothing in the file supplies an answer.
18 / 20
Key takeaways
- Labeled examples include explicit targets; unlabeled examples do not, and the model must then discover patterns such as groupings from the inputs.
- Structured data follows tables or fixed schemas, while unstructured data covers free-form text, images, audio, and video.
- Tabular data uses rows for samples and columns for features, and a target column is what makes such a table labeled.
- Time-series observations carry timestamps and chronological order, and a forecasting table needs at least an item identifier, a point-in-time target, and a timestamp.
- Text and image name modalities, not label state or structure, so a dataset is described on all four axes at once.
19 / 20
Quick check

Which summary keeps the four axes straight?
1. AA caption makes a photograph text data, and a free-form field makes a CSV unstructured media
  
  A caption is a label on an image, and one free-form field does not remove the fixed schema around it.
2. BA target column can label a table, a fixed schema keeps a CSV structured, timestamps make it a time series, and text and image are modalities
  
  Right. Each clause stays on its own axis: label state, structure, time, and modality.
3. CA timestamp column supplies the target, structure decides whether answers are correct, and modality decides whether rows exist
  
  Timestamps record time rather than supply answers, structure is about schema rather than correctness, and modality does not determine whether the data has rows.
20 / 20
8 quick checks · then the test

In the app, finishing the quick checks opens this lesson’s 10-question test, and the ones you miss come back exactly when you’re about to forget them.

Take this lesson’s test → Next lesson: Machine Learning Methods Get the app

← Previous Keep your progress · Get the app Next card →

The whole course, on your phone

Lessons you can read, audio you can listen to on the way to work, and practice that remembers what you got wrong.

Get the app

Data Types for AI Models

Labels: does the example come with its answer?

Structure: does the data follow a fixed schema?

Modality: what kind of content is it?

Keep your progress in the app

Time: are the observations ordered in time?

Four questions, four independent answers

Worked example: a forecasting table

Worked example: photos, emails, and a CSV

Key takeaways

8 quick checks · then the test

The whole course, on your phone