What Is Fine-Tuning?

Fine-tuning is the process of taking an existing pretrained model and training it further to improve its performance on specialized tasks or make its outputs more consistent. It builds on what a model already knows, using a new set of training samples to refine behavior.

There are many approaches to fine-tuning – some update all model parameters (the numerical values used to produce responses), others target a smaller set, such as those in an adapter. Fine-tuning can use supervised examples of the target behavior, or preferences between alternative responses, or external reward signals to guide parameter updates. The choice can influence training cost, infrastructure requirements, and how candidate models are evaluated. Thus the right approach depends on the goal, data availability, model, and compute resources.

Fine-tuning versus training from scratch

Training a model from scratch starts with randomly initialized parameters, so that the model must learn entirely from the training data you provide. This typically proceeds from pretraining to multiple rounds of fine-tuning. For a general-purpose language model at modern scales, pretraining in particular requires enormous datasets and substantial compute resources. Fine-tuning starts with an already-pretrained model, allowing one to build on top of a model's existing capabilities with much less data and compute.

Fine-tuning is used both during a model's initial development and after its release. Labs fine-tune modern LLMs to follow instructions and solve technical tasks before releasing them publicly. For open-weight models, third-parties can then further fine-tune the released model for their own needs.

Fine-tuning takes place both as a part of core model training, as well as by smaller teams or enterprises after a model is fully trained (this latter scenario is commonly known as custom fine-tuning). The term "fine-tuning" alone doesn’t tell you who is training, what data they are using, or whether the goal is a public model or an internal application.

A simple example

Suppose a support team needs a model to route messages to one of four queues. The model understands the message, but similar requests sometimes land in different queues. One option is supervised fine-tuning: show the model representative messages paired with the queue and response format the team wants. Then test the candidate on separate messages, including confusing boundary cases.

The test (evaluation) matters as much as the training. If the updated model routes ordinary cases better but mishandles urgent ones, the team needs to see that before relying on it. A completed training run only proves that the job completed. It doesn’t prove that the behavior improved.

Fine-tuning, prompts, and retrieval do different jobs

A prompt gives a model instructions when someone uses it. Retrieval brings selected information, such as current policy text, into that request. Fine-tuning changes the model or trains small add-on parts so additional training shapes its responses.

These approaches can be combined. A fine-tuned model may still need instructions, retrieved information, output checks, or human review. Fine-tuning isn't a live database, and it doesn’t guarantee a model will follow a rule in every case.

The boundary is practical, not absolute. Training shapes how a model responds to what it sees, while retrieval can provide information that changes after training. A system may need both. The right choice depends on what failed and how the team can verify a fix.

The short answer

Fine-tuning is additional training that adapts a pretrained model for a purpose. The exact method differs by data format: examples of desired answers, comparisons between answers, and scores for outcomes work differently. Each needs a test suited to its goal.

For a team considering its own adaptation (custom fine-tuning), the next question is whether the failure calls for training at all. If it does, the next decision is which training approach fits the evidence.

Try this in Oumi

Take one task from this definition into Oumi. Describe the behavior you want to improve, define how you’ll measure it, and compare the current model with an updated model on the same test examples, kept separate from training. Before you start, confirm that Oumi supports the method and model for your workflow.

Where to go next

• What Is Custom Fine-Tuning? Explains what changes when a team adapts a model for its own task.
• Fine-Tuning Methods: Which One Is Right for You? Explains the different ways to train and adapt a model.