Model Adaptation through Finetuning

Imagine two people who must each learn a new skill. One starts with no background at all. The other has already spent years learning related material. Most people would expect the second person to get up to speed faster and with less effort. The idea behind finetuning follows the same intuition, but it is applied to a machine learning model rather than a person.

In this article you will follow what happens to a model's weights when finetuning is applied. You will trace the weights from an earlier training run into a new one, then compare finetuning with training from scratch and look at why it usually costs less.

The key word in that definition is continuing. Finetuning does not describe building a model from nothing. It describes taking a model that already has a set of weights from an earlier round of training and resuming training from that point. The weights carry forward the results of the earlier process, and the new training adjusts them further.

The starting point of finetuning is the output of an earlier training process. The source calls that earlier process pre-training. Pre-training produces a set of weights, and those weights are what finetuning picks up. The earlier training carries its learning forward into the new run, so the new run does not have to begin from zero.

Read the diagram from left to right. The earlier training run is what produces the existing weights. Finetuning does not go back to the start of that process. It takes the existing weights as they are and continues training from them. Only the later stages of the flow are new work.

To see what finetuning changes, follow the model's weights across the two stages. Before finetuning, the model is a previously trained model, and its weights are the product of that earlier training. Finetuning then continues training from those same weights. The result is a finetuned model whose weights have been further adjusted by the continued training.

A side-by-side view separates the two approaches. A model that is trained for the first time starts without prior weights from an earlier round of learning. A model that is finetuned starts from weights that already exist. The diagram below shows how the starting point, the training steps, and the outcome differ between the two.

Reading down each column shows the contrast. Both columns end with a model, but the finetuning column begins with weights that an earlier process already produced, so its training continues rather than starts over.

The source states that finetuning typically requires fewer resources, such as data and compute. The reason is that the model already has certain knowledge from pre-training. Pre-training is the earlier training that produced the starting weights, so part of the learning has already been done before finetuning begins.

Two resources are named in the source. Data refers to the examples the model learns from. Compute refers to the processing effort needed to train. Because the model does not need to learn everything again, finetuning typically draws less on both of these resources than training a model from its beginning would.