In plain words
Training a large AI model from scratch takes trillions of words and huge amounts of computing. Almost nobody needs to do that. Instead, they start from a model that already knows language well and give it extra lessons.
That extra lesson is fine-tuning. We show the model a few hundred or a few thousand examples of exactly what we want: replies in our company's tone, medical notes in a set format, questions answered in Marathi. The model's settings are nudged slightly so that answers like these become more likely.
The model keeps almost everything it knew. It just picks up new habits. That makes fine-tuning far cheaper and quicker than starting over.
Fine-tuning is also how every chatbot is made: a raw model is fine-tuned to follow instructions and hold a conversation.
Everyday example
A new chef in a restaurant kitchen
A restaurant hires a trained chef. Nobody teaches her to hold a knife or make a gravy; she learned that over years. In her first week she cooks the house menu a few dozen times, learning this kitchen's spice levels and plating. Her skills stay the same; her habits shift to match the house. That first week is fine-tuning.
Where the example breaks: A chef can also learn new facts that week, like where the saffron is kept. Fine-tuning is good at teaching style, format and behaviour, but it is an unreliable way to add new facts. For that, giving the model documents at question time usually works better.
How it actually works
Fine-tuning is the same training process as pre-training, started from a trained model instead of random numbers, with far less data.
Prepare examples
Collect input and ideal output pairs, such as a customer question and the perfect reply. Quality matters more than quantity: a 2023 study called LIMA got strong results from just 1,000 carefully chosen examples.
Continue training
Run the model on the examples, measure how far its answers are from the ideal ones with a loss function, and adjust its parameters by small steps. A low learning rate keeps it from overwriting what it already knows.
Full or light
Full fine-tuning updates every parameter, which needs a lot of memory. Lighter methods like LoRA freeze the model and train a small add-on instead, often under 1% of the size, which makes fine-tuning possible on a single GPU.
Kinds of fine-tuning
Supervised fine-tuning learns from correct examples. Instruction tuning is supervised fine-tuning on instructions and answers. Preference methods such as RLHF and DPO learn from pairs of better and worse answers.
Watch for side effects
Too much training on a narrow set can make a model forget general skills or copy its examples too closely (overfitting). Teams test the tuned model on held-out examples before using it.
- Popularised for language
- 2018: ULMFiT (Jeremy Howard and Sebastian Ruder) and OpenAI's first GPT both pre-trained, then fine-tuned
- Small data, big effect
- ULMFiT matched models trained on 100 times more data using only 100 labelled examples
- Where it happens now
- Mostly on open-weight models such as Llama, Gemma and Mistral; as of October 2026 OpenAI says it is winding down its public fine-tuning service
- Close cousin
- Transfer learning: reusing knowledge learned on one task for another
Questions people ask
What is the difference between fine-tuning and prompting?
Prompting changes what we tell the model each time; the model itself stays the same. Fine-tuning changes the model's settings permanently. Try prompting first: it is free and instant. Fine-tune when we need the same behaviour thousands of times or cannot fit enough examples in a prompt.
Is fine-tuning better than RAG?
They solve different problems. RAG fetches documents at question time, so it suits fresh or private facts that change. Fine-tuning shapes behaviour, tone and format. Many products use both: RAG for knowledge, fine-tuning for style.
How much data does fine-tuning need?
Often far less than people expect: from a few dozen to a few thousand good examples, depending on the task. Clean, consistent examples matter more than volume. Teaching a new format needs fewer examples than teaching a hard skill.
Can anyone fine-tune ChatGPT or Claude?
Options change often. As of October 2026, OpenAI says it is winding down its fine-tuning platform, which is no longer open to new users. Most hands-on fine-tuning today happens on open-weight models that anyone can download and train.
What is the difference between pre-training and fine-tuning?
Pre-training is the huge first stage where a model learns language from trillions of tokens. Fine-tuning is the small second stage that adapts it to a task or behaviour. The first builds knowledge; the second shapes how it is used.