What Is Fine-Tuning? How Teams Adapt Models to Their Data

Fine-tuning trains a pre-trained model further on your examples so it matches your task, tone and format. How it works, the methods, and when to use it.

By Galactic News Editorial · Published 2026-11-09T22:29:14+00:00 · Updated 2026-11-09T22:29:14+00:00

What Is Fine-Tuning? How Teams Adapt Models to Their Data

Fine-tuning takes a pre-trained model and trains it further on examples of the behaviour you want. The base model already knows language in general. Fine-tuning teaches it your task in particular: your tone, your format, your categories, your edge cases. The result is a custom model that answers more reliably in your style. This guide explains how it works, when it helps, and what has changed about the main platforms in 2026.

What is fine tuning, in plain terms

Think of a base model as a strong generalist. It can write, summarise, and classify. But it does not know your house style, your product names, or your support categories. Fine-tuning closes that gap with examples. You show the model pairs of inputs and the outputs you want. Training adjusts the model's weights so its answers move closer to your examples. The tuned model stays private to your account. It does not change the base model for anyone else. Fine-tuning teaches behaviour, not new facts. It will not reliably add knowledge the base model never saw. For new facts, retrieval or prompting usually works better.

How fine-tuning works in four steps

OpenAI's documentation describes supervised fine-tuning in four major parts. The same shape applies on other platforms, including Google's Gemini tuning.

1. Build a training dataset. Collect examples that show what “good” looks like. Use realistic prompts and the exact outputs you want. OpenAI recommends starting with 50 well-crafted demonstrations. The minimum is 10 examples.
2. Format and upload the data. Training data uses JSONL format, with one complete JSON structure on every line. For chat models, each line holds a conversation: a user message and the desired assistant reply.
3. Create a fine-tuning job. Pick the base model and the training file, then start the job. Small jobs finish in minutes. Large ones can take hours.
4. Evaluate the result. Test the tuned model against the base model on held-out examples. OpenAI's advice is blunt: only invest in fine-tuning after setting up evals. Without a test set, you cannot tell whether tuning helped.

Each full pass over the dataset is an epoch. Checkpoints are saved at the end of each epoch. If the model improves early and then starts memorising the data, an earlier checkpoint may perform better. That failure mode is called overfitting.

Three methods teams use

Supervised fine-tuning is the most common method. You provide input-output examples and the model learns to reproduce them. It fits classification, formatting, translation, and style matching. Direct preference optimisation is the second method. You provide a prompt with two responses: one preferred, one not. The model learns to favour the preferred kind. It fits tone and quality judgments that are hard to write as single correct answers. Reinforcement fine-tuning is the third method. It trains a reasoning model against a grader you define. The model samples answers, the grader scores them, and training favours high-scoring answers. It fits tasks with clear right answers, such as maths or code. Google's docs also describe LoRA, a parameter-efficient mode that adjusts only a subset of parameters. LoRA costs less and needs less data than full fine-tuning, while full fine-tuning has higher quality potential.

When fine-tuning helps, and when it does not

Fine-tuning helps when the task is narrow and stable. Good cases include matching a strict output format, classifying text into your own labels, fixing repeated instruction failures, and cutting prompt length for a task you run often. It helps less when the task changes weekly, when you have only a handful of examples, or when the model needs facts it was never trained on. Try better prompting first. Prompting is faster and cheaper to change. If a tuned prompt still fails the same way on the same cases, that is the signal to gather examples and fine-tune. Keep your data clean and consistent. Contradictory examples teach contradictory behaviour. Hold out a test set before training starts, and never train on it.

A note on OpenAI's fine-tuning platform

Check platform status before planning a project. OpenAI is winding down its fine-tuning platform. The platform is no longer accessible to new users. Eligible existing customers can create training jobs until January 6, 2027, unless the base model shuts down earlier. Fine-tuned models stay available for inference until their base model shuts down. Google still documents supervised tuning for Gemini models, including Gemini 2.5 Pro and Gemini 2.5 Flash. Teams starting fresh should check which providers still accept new tuning customers.

Bottom line

Fine-tuning is a small, focused training run that adapts a general model to one job. Start with 50 good examples, set up evals first, and test against the base model. It is the right tool for stable, narrow tasks. It is the wrong tool for teaching new facts or fixing a vague prompt.