Adapt a model when adaptation is the right answer, and say when it is not
Fine-tuning teaches a model form, not facts; we baseline against prompting and retrieval first, then adapt only where a real gap survives.
Method
If the complaint is that the model does not know last quarter's policy, retrieval is cheaper and easier to audit. If it will not hold your schema, your tone or your language, that is a weights problem. We settle which before a GPU hour is booked.
Adaptation
Prompting, retrieval, LoRA, full fine-tuning, preference optimisation, continued pre-training. Cost and lock-in rise at each rung, so we stop at the first that clears your bar.
Efficiency
<1%
Trainable parameters under a typical LoRA configuration
Memory
3–4×
Typical reduction from 4-bit QLoRA against bf16 full fine-tuning
Languages
Turkish and Arabic are under-represented in most base models, and it shows in the tokenizer first: the same sentence can cost two to three times the tokens it does in English.
Data
Most of an engagement is dataset work: licence review, near-duplicate removal, decontamination against every benchmark you will report, and agreement measured, not assumed.
Training
Adaptation
Data
Evaluation
Infrastructure
Who we do this for
2–3 weeks
We baseline the frozen model, build the held-out evaluation set, audit the data and report whether training is justified at all.
Baseline comparison, held-out evaluation set, written go or no-go
6–14 weeks
Dataset curation, training runs, evaluation and regression testing, delivered with the pipeline that produced them.
Curated datasets, adapters or merged checkpoints, training pipeline, model card
Ongoing
Re-training as your data and the base models move, benchmark maintenance, contamination re-checks and migration to better bases.
Re-training runs, refreshed benchmarks, migration plan
When the knowledge changes weekly, when you have fewer than a few hundred good examples, when nobody has written down what a correct answer looks like, or when you have not yet tried a serious prompt and a decent retrieval pipeline. A fine-tune also pins you to a base model that will be superseded.
For style, format or a narrow classification boundary, a few thousand carefully written examples usually beat a few hundred thousand scraped ones. Preference optimisation needs paired comparisons. Continued pre-training for a language is a different order of magnitude, measured in billions of tokens.
No. It makes the model more like your data. Reasoning ability comes from the base model and its pre-training compute. What adaptation reliably buys is consistency, format adherence, domain vocabulary, language coverage and an equivalent result from a smaller, cheaper model. For raw capability, change base models.
Send us the constraint you are actually up against — budget, latency, regulator, deadline.
