Got it — the five additional formats are TSV, XLIFF 1.2, XLIFF 2.0, bilingual Word (.docx), and Excel (.xlsx). Here’s the rewritten copy:

---

Train your own AI with Fine-Tuning

How your existing translation data trains a personalised AI model

What is Fine-Tuning?

Fine-tuning is the process of taking a general-purpose AI language model and retraining it on a specific set of examples, so that it learns your vocabulary, style, and subject matter. The result is a model that does not just translate — it translates the way you do.

General AI models are trained on vast amounts of text from across the internet. They are capable and broad, but they have no knowledge of your preferred terminology, your clients’ house style, or the specialist language of your particular field. Fine-tuning closes that gap.

What Training Data Can I Use?

The Fine-Tuner accepts aligned bilingual data in the formats that professional translators already work with every day:

• TMX — the Translation Memory eXchange format exported by SDL Trados, memoQ, OmegaT, Wordfast, and virtually every other CAT tool
• XLIFF 1.2 and XLIFF 2.0 — the industry-standard XML localisation format used in software and content localisation workflows
• TSV — tab-separated files with column headings identifying the source and target language
• Bilingual Word documents (.docx) — a two-column table with language headings, as commonly used for review and handoff workflows
• Excel spreadsheets (.xlsx) — bilingual data across one or more sheets, with column headings identifying each language

If you have been working as a translator for any length of time, you almost certainly have this data already — in your CAT tool’s translation memory, your localisation project archives, or your client delivery files. That accumulated body of work is the foundation of fine-tuning.

How the Fine-Tuning Process Works

Local AI Translator’s Fine-Tuner reads your source file, extracts every aligned segment pair, and uses it to train a personalised AI translation model. The process runs entirely on your own machine or a local GPU server, with no data sent to the cloud at any stage. Here is what happens at each step:

1. Data conversion. The Fine-Tuner parses your file — regardless of format — and extracts every source-target segment pair. These are reformatted into JSONL (JSON Lines), the training data format expected by the AI framework.
2. LoRA fine-tuning. Using a technique called Low-Rank Adaptation (LoRA), the base model is fine-tuned on your segment pairs. LoRA is efficient and practical — it modifies only a small set of the model’s parameters rather than retraining the entire network, which means it can run on a consumer-grade GPU with as little as 6 GB of VRAM.
3. GGUF conversion. Once training is complete, the fine-tuned model is converted to GGUF format — the compressed, quantised format used by Ollama for fast, efficient local inference.
4. Ollama registration. The converted model is automatically registered with Ollama under the name you choose. From that point on, it is available to Local AI Translator just like any other model — but it has been trained on your own work.

The entire pipeline is managed through a single interface. No command-line knowledge is required.

How This Helps Translators

Terminology consistency

Every translator builds up preferred terminology over years of work. Fine-tuning encodes that preference directly into the model. If your translation data consistently renders a particular technical term one way, the fine-tuned model will do the same — without you having to correct it each time.

Subject-matter specialisation

A model fine-tuned on a medical dataset will produce different output from one fine-tuned on a legal or financial one. Rather than using a single general model for everything, you can train separate models for each of your specialist domains and select the appropriate one for each job.

House style and register

Clients often have strict requirements around tone, register, and preferred phrasing. If your data reflects years of work for a particular client, a model fine-tuned on that material will naturally reproduce their house style — reducing post-editing time significantly.

Faster post-editing

Because the fine-tuned model has learned from translations you have already approved, its suggestions are much more likely to be acceptable without revision. This is particularly valuable for high-volume, repetitive content where a general model would require constant correction.

Complete privacy

Your translation data contains confidential client material, proprietary terminology, and commercially sensitive content. Because fine-tuning runs entirely on your own hardware with no internet connection required, that data never leaves your machine. There is no upload to a third-party service, no terms-of-service clause about training data, and no risk of your files being used to train someone else’s model.

A model that improves with your career

As your data grows with each new project, you can retrain the model to incorporate your latest work. The more you translate, the more precisely the model reflects your current style and terminology — it grows with you.

What You Need to Get Started

• Bilingual translation data in any supported format (TMX, XLIFF, TSV, Word, or Excel)
• A machine with a compatible NVIDIA GPU (6 GB VRAM minimum for 7B models)
• Local AI Translator with the Fine-Tuner component installed
• Ollama running locally or on a networked GPU server