AI Trend Research

← AI glossary

AI glossary · Basic · Language models

Fine-tuning

Fine-tuning is additional training that adapts an AI model, already trained broadly, to a specific field or task.

About this term
Level
Basic · Technical terms you often meet in AI news
Added
2026-10-07
In plain words It is like taking a new employee who has completed general education and giving them separate training on the company's ways of working and technical terms. It brings them up to speed for the job much faster and at far lower cost than training someone from scratch.

How it works

First, you prepare a 'base model' trained on a massive amount of data. You then train it further on data from the field you want, such as legal documents, customer support records, or pairs of questions and model answers. This adjusts the model's parameters slightly so it better follows that field's tone, knowledge and formats. Methods that cut costs by changing only some of the parameters instead of all of them are also widely used.

Fine-tuning after pretrainingMassive general data → Pretraining; Pretraining → Base model; Base model → Fine-tuning(Starting point); Domain data → Fine-tuning; Fine-tuning → Specialized model① Pretraining · large and expensive · once② Fine-tuning · small and cheap · per useMassive generaldataPretrainingBase modelDomain dataLegal · supportrecords · Q&AFine-tuningAdjust all or onlysomeSpecialized modelDomain tone · format ·knowledgeStarting point
Fine-tuning after pretraining Expensive pretraining is done once, and the resulting base model is further trained with small amounts of data for each use. Methods that change only some parameters instead of all of them are also widely used.

Why it matters

Building a model from scratch requires enormous amounts of data and computing resources. Fine-tuning lets you adapt an existing model to the use you want with fewer resources while keeping its capabilities. It is a common way for companies to build AI specialized for their own work.

Things to keep in mind

For reflecting the latest information or frequently changing material, RAG is often a better fit than fine-tuning. Many assessments find fine-tuning more effective for teaching tone, format or a particular way of doing a task than for adding knowledge. Also, training too heavily on narrow data can weaken other abilities the model originally had.

Example

A company building a customer service chatbot can, instead of using a general-purpose LLM as is, fine-tune it on its own support records and response guidelines so that it answers in the company's tone and in line with its policies. One reason open-weight models are popular is that they are easy to fine-tune directly in this way.

Related terms

LLM RLHF RAG Open weights Distillation Parameter