What Is Fine-Tuning?

What Is Fine-Tuning?

Fine-tuning takes a pre-trained model and continues training it on a curated dataset relevant to your specific use case. The result is a model whose weights have been adjusted to perform better on your task, in your style, with your vocabulary. Unlike prompting — which provides guidance at inference time — fine-tuning changes the model itself, baking in patterns that activate without any prompting at all. A fine-tuned model might produce perfectly formatted legal clauses without being told to, because it learned that format from thousands of examples.

When to Fine-Tune vs Prompt

Fine-tuning is worth the investment when: (1) you need consistent style or format that would require very long, repeated system prompts to maintain; (2) you have hundreds or thousands of domain-specific examples of correct behaviour; (3) you need the model to use proprietary terminology, abbreviations, or conventions that don't appear in general training data; or (4) you need to reduce token costs by eliminating verbose system prompts. If prompting with good examples can already hit your quality target, start there — fine-tuning has a data preparation and compute cost that's hard to justify for marginal gains.

Cost, Effort, and Overfitting Risk

Preparing a fine-tuning dataset requires careful curation: each example must demonstrate exactly the behaviour you want, labelled consistently, in the format the fine-tuning API expects. Poor quality data produces poor quality fine-tunes — the model learns your errors as faithfully as your correct patterns. Overfitting is a real risk: a model fine-tuned on a narrow dataset can become brittle, performing well on in-distribution examples but poorly on anything outside the training distribution. Regularisation techniques and validation sets help, but this requires ML engineering expertise.

LoRA and PEFT Methods

Full fine-tuning updates all of a model's billions of parameters, which is computationally expensive and requires significant GPU memory. Parameter-Efficient Fine-Tuning (PEFT) methods, especially LoRA (Low-Rank Adaptation), solve this by only updating a small number of additional parameters layered on top of the frozen base model. LoRA fine-tunes can run on consumer hardware and produce results competitive with full fine-tuning for most tasks. Most cloud fine-tuning APIs (OpenAI, Together, Replicate) use LoRA or similar approaches under the hood.