AI Specialization Program

WEEK 2 Lesson 1

Title: Understanding Datasets & Data Splits.

Theme: Engineering Data Before Models

Core Principle: Bad Data = Bad Model

Lesson Objective

By the end of this lesson, students will:

✔ Understand how datasets are structured

✔ Differentiate features vs labels

✔ Apply proper train/validation/test splits

✔ Detect and avoid data leakage

✔ Think with a dataset documentation mindset

This lesson builds the foundation of reliable AI systems.

1. What Is a Dataset?

A dataset is a structured collection of data used to train or evaluate a model.

In AI systems, datasets usually contain:

° Rows → observations (examples)

° Columns → features (attributes)

° Optional column → label/target (what we want to predict)

Example 1: Structured Dataset (Tabular)

See the data tables below.

Examples:

Loan approval dataset

Hospital patient dataset

Agricultural yield dataset

School performance dataset

Each row = one instance

Each column = one feature.

 Example 2. Unstructured Dataset (See below illustrations)

Examples of unstructured dataset:

a. Text (news articles, SMS data)

b. Images (crop disease photos)

c. Audio (local language recordings)

d. Video (traffic footage)

Unstructured data must often be transformed before modeling.

2. Features vs Labels

This is one of the most important concepts in AI.

🔹 Features (X)

Inputs used by the model.

Examples:

Age

Income

Temperature

Text message content

Image pixels

🔹 Label (y)

The output we want to predict.

Examples:

Approved / Not approved

Disease type

Sentiment (Positive/Negative)

Crop healthy vs diseased

Simple Example:

Age    |   Income    |   Loan Approved

35      |   50000     |     Yes

Features → Age, Income

Label → Loan Approved

Without a clear label, supervised learning cannot happen.

3. Train / Validation / Test Split

This is where many beginners fail.

Why split data?

If you train and test on the same data:

Your model memorizes.

You get fake performance.

You deploy failure.

Proper Split Structure below.

No

Typical ratios:

70% → Training

15% → Validation

15% → Test

🔹 Training Set

Used to train the model.

🔹 Validation Set

Used to tune hyperparameters and adjust model choices.

🔹 Test Set

Used once. Only at the end. For final unbiased evaluation.

4. What Is Data Leakage?

Data leakage happens when:

Information from the future leaks into training.

This gives artificially high performance.

 Example of Leakage

Imagine predicting loan default:

If your dataset includes:

“Loan Repaid Status After 6 Months”

And you use that as a feature…

You are cheating.

The model sees the answer indirectly.

Common Leakage Sources:

Mixing train and test data

Scaling entire dataset before splitting

Using post-outcome features

Time-series data shuffled incorrectly.

Leakage destroys trust.

Engineers must guard against it.

5. Dataset Documentation Mindset

Professional AI engineers document datasets.

This prevents misuse and bias.

A good dataset document includes:

Source of data

Date collected

Number of samples

Number of features

Label definition

Known limitations

Known biases

Missing value summary

This is inspired by practices used in major AI research institutions like Google Research and OpenAI.

Serious systems require serious documentation.

African Context Awareness

When working with African datasets:

Language diversity matters

Infrastructure gaps affect data quality

Informal sector data may be inconsistent

Historical bias may exist in financial or health datasets

Data is not neutral.

Engineers must think ethically.

 Mini Practical Exercise

Students must:

1. Download a public CSV dataset.

2. Identify:

    ° Features

    ° Label

3. Perform:

     ° 70/15/15 split

4. Write 5 bullet points documenting the dataset.

Identify at least one possible leakage risk.

5. Deliverable: Short technical summary.

 Lesson 1 Outcome

By the end of this lesson, students:

✔ Understand structured vs unstructured data

✔ Identify features and labels

✔ Perform proper dataset split

✔ Detect leakage risks

✔ Think like responsible AI engineers

Engineering Principle of the Week

“The model is only as good as the data discipline behind it.”

Next:

🔷 Week 2 — Lesson 2: Pandas for Data Loading & Exploration

Powered by Soft AI Africa | Training the Next Generation of AI Leaders in Africa.