Service · Fine-tuning

Where prompting alone stops working.

Most teams should not fine-tune. Better prompting, better retrieval, and better tool design solve 80% of cases people ask us to fine-tune for. The 20% that genuinely need it - domain-specific terminology, your company's voice, output formats the base model can't reliably produce, latency requirements that need a smaller model - are where this service exists.

Last updated:

The problem

Almost what you want, consistently off.

You've optimised the prompt. You've added retrieval. You're running few-shot examples. The output is almost what you want, but consistently off in a way no prompt-engineering trick fixes. Or you need a 7B model because latency budget is sub-second and the 70B base is too slow.

That's where fine-tuning earns its keep - and where most teams should not start, because better prompting and better retrieval get them most of the way there for a fraction of the work.

What's included

Six things every fine-tune engagement gets.

  • Pre-fine-tune diagnostic

    First, we try to make your problem go away with prompting + retrieval. About 30% of teams who came to us for fine-tuning did not need it once we ran that pass.

  • Data preparation

    Extracting, deduplicating, formatting your training data. This is 70% of the work - we will not pretend otherwise.

  • Base-model selection

    Which model fits your size / latency / quality / cost target. Qwen, Gemma, Llama, GLM, Kimi, Nemotron, DeepSeek, Mistral on the open-weights side; Claude, Gemini, GPT, Grok via the provider fine-tune APIs.

  • Training run

    On our hardware infrastructure for open-weights, or via the provider's fine-tune API for cloud models.

  • Evaluation

    Golden-set tests, A/B against the base model, regression tests for quality drift.

  • Deployment + retraining cadence

    Onto your local-LLM hosting or the provider API. Schedule for keeping the model fresh as your data evolves.

Where fine-tuning fits

Last lever, not first.

Most accuracy gains come from better prompting and retrieval. Fine-tuning is the lever you reach for when those have plateaued and the cost-benefit lines up.

Better prompting + retrieval
First lever · cheaper · faster to iterate
  • No training data preparation needed
  • Updates instantly when you change the prompt
  • Works across model versions automatically
  • Solves ~80% of "the model is wrong" cases
  • Cheap to maintain
Fine-tuning
Last lever · for the remaining 20%
  • Domain vocabulary the base model never saw enough of
  • Output formats prompting can't reliably enforce
  • Latency budget that demands a smaller model
  • High-volume generation where consistency compounds
  • Distillation for inference-cost cuts
When fine-tuning is the right answer

Five situations that genuinely justify it.

  • Domain-specific terminology

    Legal documents in less-represented languages, niche industrial codes, sector-specific abbreviations and conventions the base model handles inconsistently.

  • Tone-of-voice at scale

    Customer-facing copy at 1,000+ generated emails per day. Prompting works for 50/day; fine-tuning works for 1,000.

  • Structured outputs

    The base model can't reliably produce your specific JSON shape, your specific Markdown layout, your sector's report format.

  • Latency requirements

    Sub-second budget that needs a smaller model. A fine-tuned 7B can match a 70B base on a narrow task.

  • Distillation

    Teaching a smaller, cheaper model what a larger one knows about your domain. Cuts inference cost 4-10×.

Case study

A vision-language fine-tune that recognises a decade of stage footage.

A vision-language model fine-tuned on a global events company's flagship events - recurring speakers, recurring stage layouts, recurring branding. Lifts recognition accuracy enough to make 119 TB of historical footage searchable by plain-English prompt. Hosted on a Mac Studio 512 GB; data stays inside the client's GDPR perimeter.

119 TB
Historical footage indexed frame-by-frame
512 GB
Unified memory · Mac Studio
GDPR
Data-in-transit consent boundary
Real-time
New uploads indexed in minutes
FAQ

Three things people always ask.

Will fine-tuning reveal our training data?
No, when done correctly. We use techniques (LoRA, careful data sanitisation) that minimise leakage. The trained weights are yours; the original data stays with you.
Can we update it later?
Yes. Either retrain from scratch with new data, or layer additional fine-tunes (LoRA stacking) for cheaper incremental updates.
What if the base model gets a major upgrade?
A refresh is typically 30-50% of the original engineering cost. Worth budgeting for once a year on actively-evolving bases.
AI that already runs

Tell us where prompting hit its ceiling.

30-min diagnostic call walks through your case and shows whether fine-tuning is the right next lever, or whether prompting and retrieval still have headroom. Free; written summary either way.

Live · powered by Gemma-4 · running on our hardware in Sofia