Epochs

An epoch is one complete pass through the training data. If a model trains for 3 epochs, it has seen the full dataset 3 times.

Epoch Playground

Run one full pass at a time

Press Run Epoch to send the model through all 5 examples once. Each pass nudges the loss down and the model gets a little more familiar with the patterns.

0 Epochs

Epoch 0
Model has not seen the data yet.

Training loss High loss

Low (good)High (bad)

One epoch = the model has seen every example once.

Why Multiple Epochs?

Seeing the data more than once

A model often needs to see the data more than once. Repeated passes can help it learn patterns better.

First pass

Learns basic patterns.

Second pass

Strengthens those patterns.

Third pass

Improves confidence.

Too Few vs Too Many Epochs

Finding the sweet spot

More epochs help, up to a point. Too few and the model barely learns; too many and it starts memorising. There is no single magic number, it depends on the data and model.

Too Few Epochs

The model may not learn enough.

Good Number of Epochs

The model learns useful patterns.

Too Many Epochs

The model may memorize the training data instead of learning general patterns.

The best number of epochs depends on the data and model, teams find it by testing the model on new examples, not a fixed rule.

🔑

Key takeaways

• One epoch means one full pass through the training data.
• Multiple epochs help the model learn patterns better.
• Too few epochs may undertrain the model.
• Too many epochs may cause memorization.
• Epochs are part of training, not inference.

Summary

An epoch tells us how many times the model has gone through the full training dataset. More epochs can help learning, but too many can make the model memorize instead of generalize.

Previous: Training Loss Next: Inference