epoch

advancedwidespread

An epoch is a full learning cycle: the machine learning model goes through all examples used for training once.

In plain terms

Imagine you are preparing for an exam from a thick textbook. When you read it from beginning to end once, that is one round. In machine learning, this round is called an epoch: the model has seen the entire training material, meaning every single example in the training data.

One reading is rarely enough. Just as you review material multiple times, models are usually trained over several epochs, and they typically become slightly more accurate in each round. The number of epochs is therefore an important training setting: it tells the machine how many times to repeat the lesson.

More is not always better. Someone who reads the textbook for the twentieth time no longer understands it but memorizes it, and gets stuck on a slightly differently phrased question in the exam. The same happens with the model: after too many epochs, it recites the training examples by heart but performs worse in new situations.

How it works

Training data is split into smaller batches. First, the entire dataset usually does not fit in the machine's memory at once; second, the model learns faster and better in many small steps.
The model receives batches one by one: for each, it guesses, the error magnitude is checked, and its internal settings are subtly adjusted accordingly.
When the last batch has been processed, an epoch ends. At this point, the model is usually tested on examples it did not see during training.
Before the next epoch, the examples are mostly shuffled so that the model does not learn the order, then the next round begins.
Training is stopped either after a predefined number of epochs or when performance on unseen examples no longer improves or begins to decline.

An everyday example

In an article about AI or a model description, you might read that it was 'trained for three epochs on the company's own customer emails.' This means the model went through all the emails three times before being put to work.

Who is affected

developerIt must decide how many epochs the training should run, because this determines both the model's quality and the computational cost. → Code

entrepreneur, executiveIf a company tunes a model on its own data, the number of epochs appears directly on the invoice and affects turnaround time.

parent, student, teacherAnyone learning or teaching machine learning encounters this setting immediately in their first practical tasks. → Education & Family

A common misconception

MythThe more epochs a model is trained for, the smarter it becomes.

In realityBeyond a certain point, the model merely memorizes training examples, and its performance worsens on new, unknown cases. Good training stops at the right moment.

Limits and risks

The number of epochs alone says nothing about model quality: ten rounds on a small, poor dataset give a worse result than one round on a large, good dataset.
There is no universally ideal number; it must be tested for each task and dataset.
Every additional epoch costs computing time, energy, and money.
The largest language models are often run through the massive training data only roughly once, so for them, the number of steps or processed tokens is tracked instead.

Similar, but not the same

iteration (training step) – An iteration is the processing of one batch and one update, while an epoch consists of as many iterations as the number of batches the data was split into.
batch (adag) – A batch is a small group of examples the model receives at once, while an epoch is one full pass over all the data.
episode – In reinforcement learning, an episode is a single attempt from start to finish, such as a played game, not a full set of training data.

Related concepts

Broader

machine learning, model training, Hyperparameter

Narrower

iteration, batch

On the edge

overfitting, early stopping, learning rate, training data, fine-tuning

A brief history

The word originates from the Greek 'epokhé', which originally meant a pause or a fixed point in time. Later, it came to denote the starting point of an era, and then the era or period itself. Researchers in neural networks adopted this meaning of 'period' to refer to each complete cycle of training, and the expression became part of the field's basic vocabulary.

Where things stand

By the end of 2026, epoch is a fundamental concept in machine learning, appearing in the settings of almost every widely used training tool and fine-tuning service. For the largest models, the duration of training is increasingly described alongside it by the number of training steps and processed tokens.

Status: 2026 November

Also known as

epoch · epocha · training epoch · training round · training loop · training epoch · epochs · number of epochs

Sources

Google Machine Learning Glossary – epoch

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.