LLM fine-tuning services, from audit to production.
Ertas is an applied AI lab that provides LLM fine-tuning services. We train an open model on your organisation's own data for one task, test it against your quality bar, and run it Ertas-managed, in your own cloud or on-premise. We work with organisations anywhere. If you're in Australia and the data has to stay in the country, your own cloud account in an Australian region or your own servers keep it there.
What does an LLM fine-tuning engagement include?
Five stages, and you decide whether to continue after each one. The proof of concept settles the right model size and the cost crossover before the production build starts.
Ongoing improvement runs through the adaptive model endpoint. The endpoint is currently invitation only, and we are developing a more streamlined version for wider access.
- Assess
- A free audit of where your AI spend goes and which tasks repeat at volume.
- Prepare
- We turn your examples and request logs into training and test sets. Your team approves what goes in.
- Build
- A proof of concept of three to four weeks finds the right-sized model for your bar. Then we train the production model.
- Run
- We serve it behind one API: Ertas-managed, in your own cloud or on-premise.
- Improve
- New versions train on your traffic in the background and go live only after they pass your evaluations and you sign off. If a new version falls short, the current model stays in service.
When is fine-tuning the right choice?
Fine-tuning fits a task that repeats at volume in a stable format, where a frontier model already does the job and the bill or the response time is the problem. It teaches the model a behaviour: your output format, your categories, your house style.
Knowledge that changes every day belongs in retrieval, and many builds use both. Continued pretraining vs fine-tuning covers how domain knowledge enters a model. Open-ended work, or a workflow that still changes every week, usually isn't ready, and the audit will tell you.
What does LLM fine-tuning cost?
The build is priced per task, and the price depends on the task, data volumes and the deployment environment. We set it after the audit, in a written scope.
Once the model is live, usage costs come down on two fronts: a lower price per token than the large frontier models, and fewer tokens per task, since the instructions and context you repeat in every prompt are trained into the model. The cost per task calculator shows both for your own volumes.
Running costs take one of three forms, listed on our pricing page: per token for variable or growing traffic, seats (a flat monthly fee for unlimited tokens, with a concurrency cap) for steady, predictable traffic, or a dedicated GPU for the highest volumes.
How do you know the fine-tuned model is good enough?
Before training, we build a test set from your own examples and agree the quality bar with your team. Every model version is scored against it before it takes traffic.
The method follows our published research: a frozen protocol, thresholds set in advance and confidence intervals. Domain-specific LLM evaluation walks through it step by step.
Common questions.
Which models do you fine-tune?
Open models, chosen per task. The proof of concept finds the right-sized model that holds your quality bar, and it is retrained as better base models ship.
How long does LLM fine-tuning take?
After the audit and data preparation, the proof of concept runs three to four weeks. The production model follows once you approve the result.
Do we own the fine-tuned model?
The model is trained on your data and shaped by your team, and it runs on Ertas by default. Ownership and handover of the model are available with a minimum engagement commitment.
Can our own engineers fine-tune models instead?
Yes. Your ML engineers can train models themselves in Ertas Studio, and the Enterprise plan adds frontier GPUs, a custom credit allocation and dedicated support.