Most small businesses should start with good prompts and, if needed, retrieval of their own documents, not fine-tuning. Fine-tuning and LoRA become worth it when you have hundreds of high-quality examples of a repeated task and a prompt alone cannot get the style or format consistently right, for example classifying enquiries or writing product descriptions in a fixed house style.
Fine-tuning means continuing to train an existing language or image model on your own examples so it learns a pattern. Full fine-tuning updates all of a model’s weights, which is expensive. LoRA, short for low-rank adaptation, is a technique introduced by Microsoft researchers that trains a small set of extra parameters on top of a frozen model. The result is a lightweight adapter file that changes the model’s behaviour at a fraction of the cost. LoRA is widely used with open models such as Llama, Mistral and Stable Diffusion, and hosted fine-tuning services from OpenAI, Google and others offer similar results without managing hardware.
What problem does fine-tuning actually solve?
Fine-tuning teaches form, not facts. It is good at making a model follow a format, adopt a tone, or classify inputs the way your team does. It is a poor way to teach changing information like prices, stock levels or opening hours, because those change faster than you can retrain.
For facts, retrieval-augmented generation is usually better: the model looks up current documents at question time. Many businesses that think they need fine-tuning actually need a well-organised FAQ and a retrieval setup.
Do you need fine-tuning or a better prompt?
Ask three questions before spending any money:
- Have you tried a detailed prompt with five or more good examples? Few-shot prompting solves many style problems.
- Is the task repeated at least hundreds of times a month? If not, editing outputs by hand is cheaper.
- Do you have at least a few hundred clean input and output pairs that you would be happy for a model to copy?
If the answers are no, yes and no, the honest next step is to collect examples, not to train.
Which small business tasks suit fine-tuning?
Good candidates have a clear input, a clear output and a consistent standard. A furniture shop with 3,000 products might fine-tune a model to turn supplier spreadsheets into product descriptions in its own voice. A clinic might train a classifier that sorts web enquiries into appointment, price question, complaint or other, so staff answer the right ones first.
Image LoRAs suit businesses that need many visuals in one consistent style, such as illustrated social posts. They should never be trained on images you do not have rights to, and training on a real person’s face needs that person’s consent.
What data do you need?
Quality matters more than quantity. A few hundred carefully checked examples often beat thousands of inconsistent ones, because the model copies mistakes as faithfully as good work.
Preparing a training set
- Collect real examples from your own work: approved descriptions, correctly labelled enquiries, replies your best staff wrote.
- Remove personal data such as names, phone numbers and addresses unless you have a clear legal basis.
- Make the format identical across examples, usually a JSONL file with input and output fields.
- Hold back 10 to 20 percent of examples for testing; never train on them.
- Write a one-page standard describing what a good output looks like, so you can grade results.
How much does it cost and how hard is it?
There are two broad routes. Hosted fine-tuning from a model provider means uploading a file, paying per training token and then per use. It needs no hardware and little code. Self-hosted LoRA on an open model needs a GPU, either rented in the cloud or local, plus someone comfortable with tools such as Hugging Face’s PEFT library.
Prices change often, so check current pricing pages. For a small business, the bigger cost is usually staff time to prepare and check data, and the ongoing job of retraining when your standard changes.
How do you know if it worked?
Compare the fine-tuned model with your best prompt on the held-out examples. Score both against your one-page standard. If the fine-tuned version is not clearly better, keep the prompt; it is easier to change tomorrow.
For classifiers, measure accuracy per category, not overall. A model that is right 90 percent of the time but misses most complaints is a problem.
What are the risks of fine-tuning?
A fine-tuned model copies your data, including its flaws. If your past product descriptions contain outdated claims or inconsistent spelling, the model will repeat them confidently. Clean examples are not optional.
There is also a maintenance risk. When the base model is retired or updated, your fine-tune may need to be redone. Keep the training file, the test set and the scores in one folder so retraining is a routine job rather than a new project.
How Alien Road applies fine-tuning and LoRA
In a typical engagement the first deliverable is often a recommendation not to fine-tune yet. We review the client’s task, test a structured prompt with examples, and only move on if the gap is real.
When fine-tuning makes sense, we help build and clean the training set from the client’s own approved work, run a hosted fine-tune or a LoRA on an open model, and compare it with the prompt baseline on held-out data. The client receives a short report with scores, cost per thousand outputs and a retraining plan, and keeps ownership of the dataset.
Where to start this week
Choose the single text task your team repeats most. Write the one-page standard for it, then start a shared folder of approved examples, adding every good one your team produces.
In parallel, build a prompt with five of those examples and test it on ten new inputs. If the prompt already gets you most of the way, you have saved yourself a training project. If not, your folder is already becoming the dataset.
Set a review date one month out. By then you will know whether the gap between prompt output and your standard is closing on its own or whether training is really the next step.
Keep reading: AI Optimization · AI Optimization