Glossary entry

Pretraining

The broad first phase in which a model learns patterns from a very large dataset.

Definition

What it means

Pretraining optimizes a model on large and diverse collections of text, code, images, audio, or other data. It builds general capabilities before the model is adapted to specific behaviors or tasks.

Why it matters

Role in the system

It supplies the general knowledge and representations that later training stages refine.

Example

In practice

A language model predicts the next token across a large mixture of web, book, and code data.