Supervised fine-tuning
Translation and localisation
How do I fine-tune a model on our terminology and tone across languages?
Train on your own translation memory. General models translate correctly and render your product names, technical terms and register wrongly — which is what reviewers actually send back.
| Objective | Supervised fine-tuning |
|---|---|
| Dataset shape | Source and approved target pairs from your translation memory, one direction per pair. |
| Metric that decides it | Reviewer edit distance on held-out segments, plus terminology compliance measured against your glossary. |
| Cheapest starting point | SmolLM3 3B on RTX 3080$0.09 per GPU-hour |
Machine translation has been good enough for years on general text. The remaining work in a localisation pipeline is almost entirely terminology and register: does it call the feature what you call it, does it use the formal or informal second person, does it keep the placeholder syntax intact.
Your translation memory already contains the answers, approved by a human, in exactly the format a training set wants. It is the rare case where the dataset exists before the project starts.
Measure edit distance rather than a reference-similarity score. What you care about is how much a reviewer has to change, and that is directly measurable from your existing review workflow.
The mistake to avoid
Mixing directions in one training set without marking them. The model blurs the two and starts producing source-language artefacts in target-language output.
Models to start from
In order. Start at the top and move down the list only if the evaluation gate says you have to — the smallest model that clears your bar is the one you pay for on every request forever.
Before you deploy it
Whatever the metric above says in isolation, the deployment decision is comparative: is this better than what is running today? Gate the candidate against the incumbent, blinded, on held-out cases from your own traffic. A fail is a cheap answer to an expensive question.
Frequently asked questions
One model per language pair, or one for all?
One multilingual model is usually enough and far cheaper to serve. Split only when a specific pair underperforms and more data has not fixed it.
Does this replace human review?
No. It reduces how much reviewers change, which is where the cost is. Removing review is a separate decision with different risks.
Start with SmolLM3 3B
Upload your dataset, read the validation report, and forecast the run before anything is leased.