Fine-tuning is the process of training an existing AI model further on your own examples so it learns a specific style, format or task. LoRA, short for low-rank adaptation, is a cheaper way to do this: it trains a small add-on layer instead of changing the whole model. For marketers, both are worth considering only after good prompts and retrieval have been tried and found lacking.
What does fine-tuning change in a model?
A base model already knows language, grammar and a lot of general facts. Fine-tuning shows it hundreds or thousands of input and output pairs, such as a product spec and the description your team wrote for it, so the model learns to produce that kind of output by default.
It is good at teaching behaviour: tone, structure, length, labels and formatting. It is poor at teaching facts that change. If you fine-tune on this season’s prices, the model will repeat them long after they are wrong.
OpenAI offers fine-tuning for several of its models through the API, Google offers tuning in Vertex AI, and Amazon Bedrock supports customisation for selected models. Open-weight models such as Llama, Mistral and Qwen can be fine-tuned on your own hardware or through services such as Hugging Face.
How is LoRA different from full fine-tuning?
Full fine-tuning updates all of a model’s parameters, which takes a lot of GPU memory and produces a full copy of the model for every version. LoRA freezes the original weights and trains small matrices that sit alongside them. The result is an adapter file that is often a small fraction of the model’s size.
That has practical benefits. You can keep one base model and swap adapters per brand or per task. Training is faster and cheaper. QLoRA, a related method, trains on a compressed version of the model and makes tuning possible on a single strong GPU.
LoRA is also widely used for image models. Designers train LoRA adapters on Stable Diffusion or Flux to reproduce a product, a character or an illustration style consistently, which is something prompts alone struggle to do.
When should marketers use prompting, RAG or fine-tuning?
Most problems are solved by the cheapest option. Move up the list only when the previous step has been tested properly and fails.
- Prompting with examples: fixes tone and format for most tasks, costs nothing to change.
- Retrieval-augmented generation (RAG): gives the model your current documents at answer time, best for facts that change.
- Fine-tuning or LoRA: teaches a consistent behaviour across very high volume, or a style that examples in a prompt cannot capture.
- Training a model from scratch: almost never sensible for a marketing team.
A useful test question: does the problem involve knowing something or doing something? Knowing points to RAG. Doing, in the same way thousands of times, may point to fine-tuning.
Which marketing tasks are good candidates?
Good candidates are narrow, repeated and well documented, with a clear right answer or a stable house style. You also need clean examples; the model will copy every mistake in the training set.
- Classifying support tickets or search terms into your own category list.
- Writing product descriptions in a fixed format across a large catalogue.
- Producing ad copy variants that follow strict brand and legal rules.
- Generating on-brand images of a product or mascot with an image LoRA.
- Extracting structured data from supplier documents.
What does a fine-tuning project need?
Data
Collect a few hundred to a few thousand high-quality examples, depending on the task. Remove personal data, outdated claims and anything you would not want repeated. Split off a test set that the model never sees during training.
Evaluation
Decide how you will judge success before training. For classification, use accuracy on the test set. For writing, use a blind review where editors compare tuned output with prompted output without knowing which is which.
Maintenance
When the base model gets a new version, your adapter may need retraining. Budget for that, and keep the training data and settings in version control.
What are the risks?
Fine-tuned models can become narrower. A model tuned heavily on product descriptions may get worse at general instructions. Tuning can also weaken some safety behaviour, which is why hosted providers run checks on training data.
Licensing matters for open-weight models and for image LoRAs trained on photos or illustrations. Train only on material you own or have rights to use, and record where each training file came from.
How Alien Road applies fine-tuning and LoRA
Our AI optimisation team treats fine-tuning as a later step. Discovery starts by testing the task with strong prompts and, where facts matter, a retrieval setup. Only if those fall short on quality or cost at high volume do we propose tuning.
When we do, we build the training set with the client’s editors, remove personal and outdated content, and hold back a test set. Image LoRAs are trained only on product photography the client owns.
Reporting compares the tuned model with the prompted baseline on accuracy, edit time and cost per thousand outputs. If the gain is small, we recommend staying with prompts, since they are easier to maintain.
How much does it cost?
Hosted fine-tuning usually charges per training token plus a higher rate per token when you use the tuned model. LoRA on open weights costs GPU hours, which can be modest for small models. The larger cost is people: preparing data, evaluating results and retraining when things change. Price those hours in before comparing with prompting.
Keep the decision simple for stakeholders. Show the baseline, the tuned result and the yearly maintenance cost on one slide. If the tuned model wins clearly on a task the business repeats every day, the case makes itself; if the difference needs a long explanation, it is probably not worth the upkeep.
Where to start this week
Pick one repeated task where output quality is inconsistent. Write the best prompt you can with three examples and run it on fifty real inputs. Score the results. If more than a few need heavy editing, collect the edited versions as they come in; that becomes the core of a future training set, and you will know exactly what the tuned model has to beat.
Keep reading: AI Optimization · AI Optimization