← The glossary

training

To begin with

We say a model “learns”, but it doesn’t read a textbook or memorise sentences. It adjusts settings, a bit like tuning an instrument by listening to the note.

The facts

The model processes examples and produces outputs. A loss function measures the gap with the goal, backpropagation works out how each parameter affects that gap, then an optimizer adjusts the parameters. This cycle repeats over many batches of examples. Pretraining and fine-tuning are two forms of training.

Key takeaway

For a model, learning means changing its weights. Answering a question is inference: nothing is relearned.

The stories about it

The notebook opens soon.The first stories are on their way.