SoloDevMeet
← Courses

Course · AI Lab

Data Training

A famous rule of thumb in machine learning is that most of the real work isn't designing clever algorithms — it's preparing good data and training carefully. This overview covers what "training" actually computes and how to know whether a model genuinely learned something useful, illustrated before you move into the interactive modules below.

Splitting your data before you touch it

Before training begins, a dataset is split into separate portions with strict jobs. The training set is what the model actually learns from. A validation set is held back to tune decisions during development without contaminating the final judgement. A test set is touched only once, at the very end, to produce an honest, unbiased score — because the moment you start adjusting your model based on test results, that score stops being trustworthy.

Training set (~70%) — the model learns from thisValidationTestwhat the model sees and adjusts ontunes settingsduring trainingfinal honestgrade — seen onceThe test set is never touched until the very end — that is what keeps the score honest.
Training data teaches the model; validation tunes it along the way; the test set gives one honest final score.

What "training" is actually computing

During training, the model makes a prediction, compares it to the correct answer, and measures how wrong it was — that measurement is called the loss. Gradient descent is the algorithm that repeatedly nudges the model's internal settings in the direction that reduces that loss, a small step at a time, the same way you might feel your way downhill in fog by always stepping toward the steepest downward slope you can sense from where you're standing.

minimum errorstart: high errorEach step nudges the model's settings a little further downhill, guided by the gradient.
Gradient descent takes repeated small steps downhill on the error curve until the loss is minimised.

The trap that catches almost everyone: overfitting

A model that fits the training data too closely starts memorising its noise and quirks instead of learning the real underlying pattern — that's overfitting, and it performs beautifully on data it has already seen while failing on anything new. The opposite problem, underfitting, means the model never learned the pattern well enough in the first place. The goal is always the middle ground: a model that generalises to new data it has never seen.

UnderfitGood fitOverfit
Underfitting misses the pattern. Overfitting memorises noise. A good fit generalises to new data.

Continue hands-on below

The modules below let you clean messy data, split datasets correctly, and run real training code with a real console — this overview is the map; the modules are where it becomes a skill.

Continue in AI Lab

Interactive modules for Data Training

Signal Start

Working With Data

The unglamorous but critical part of ML: what makes data usable for training, how to clean it, and how to split it so your evaluation is honest.

6 guided lessons →

Pattern Builder

How to Train a Model

The "train data" course: what training actually computes, hands-on JavaScript labs for loss, gradient descent, and the learning rate — with real running code and a real console.

7 guided lessons →