MMOOZZGG.com

Glossary

What is model training?

In short

Model training is the process in which a program repeatedly compares its answers with the right ones and adjusts its internal parameters until the errors are few.

AI basics · training

MMOOZZGG.com editorsChecked: 1 min read

For a child

Remember how you learned to throw a ball into a basket? You threw, missed to the right, and next time aimed a bit to the left. Each try told you something. A computer learns in a similar way: it makes a guess, finds out how far it missed, and changes itself a little. After very many tries, it hits almost every time.

In plain words

Before training starts, the program answers at random. You give it an example, say a photo labeled “dog”, and see what it said. If the answer is “cat”, a special formula computes how far that is from the truth. It also shows which way to move each internal setting. Step by step, over thousands of examples, the error shrinks. One full pass through all the examples is called an epoch, and there are many of them. When the result on new, unfamiliar examples stops improving, you stop training.

For business

Companies rarely train a large neural network from scratch: it takes huge computing power and data. They usually choose one of three paths. They take a ready system with good instructions, connect their own documents to it through RAG, or fine-tune it on their own examples. Fine-tuning gives the right style and domain knowledge. But it costs money to prepare examples and rent GPUs, and you have to repeat it when a new base version comes out. The risk: if the examples contain errors or personal data, the system will remember them. Compare which path pays off with a calculation on a specific process, before you buy servers.

For a specialist

Training is the optimization of the parameters θ against a loss function averaged over the training set. Gradients are computed by backpropagation, which D. Rumelhart, G. Hinton and R. Williams showed for multilayer networks in Nature in 1986. Parameters are updated by stochastic gradient descent (SGD) on mini-batches or by adaptive methods. A common one, Adam (D. Kingma, J. Ba, 2014), sets the step for each parameter from estimates of the first and second moments of the gradient. Key hyperparameters: learning rate, batch size, number of epochs and regularization (weight decay, dropout). Overfitting is tracked on a validation set, and training ends early (early stopping). Large language models go through these stages: pretraining on large corpora, supervised fine-tuning on instructions, and preference alignment (RLHF, DPO). Cheaper ways to fine-tune: LoRA and other PEFT methods.

Example

A bank fine-tunes an open neural network on a thousand of its own replies to customers. After that, the assistant writes in the bank’s usual tone and names its products correctly.

Related terms

Questions and answers

How long does model training take? From minutes to months. A small model on a table of data trains in minutes on an ordinary computer.

Fine-tuning a language model takes hours or days on GPUs. A large neural network trained from scratch needs big computing clusters, and that is expensive.

How is training from scratch different from fine-tuning? Training from scratch starts with random parameters and needs a huge amount of data. Fine-tuning takes a system that is already trained and adjusts it on a small set of your own examples, so it costs much less.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us