Techbanq

AI · model adaptation

Fine-tuning is the last resort, not the first move

Changing a model's weights is the most expensive way to change its behaviour, and most enterprise cases do not need it: retrieval puts your knowledge in front of the model and prompting shapes how it answers. We start with a baseline evaluation, and tune only when the evidence says the cheaper levers have run out.

Serving Saudi Arabia from our Riyadh office and our engineering hubs in Lahore and Karachi

Market
Saudi Arabia
Written for
CTO, Head of Digital
Decision
Decide whether to fine-tune a model or solve the problem with retrieval and prompting

Enterprise delivery credentials

  • ISO

    ISO/IEC 27001

    ISO/IEC 27001 certified

Certifications & registrations

The decision

Prompting, retrieval or fine-tuning: what each lever actually changes

These are not three tiers of the same thing; they change different parts of the system and carry very different costs of ownership. The order below is the order in which they should be tried.

OptionWhat it changesWhat it needs from youCost to keep currentBest for
PromptingHow the model is instructed: tone, format, rules, examples in the prompt itself.Clear instructions and worked examples — no data pipeline at all.Edit a prompt; the change ships the same day.Shaping behaviour, format and tone.
Retrieval (RAG)What the model knows at answer time: your documents are fetched and placed in front of it per question.Well-prepared content — normalised, chunked, access-controlled.Update the documents; the model answers from the new version immediately.Answering from your own knowledge.
Fine-tuningThe weights themselves: vocabulary, style and task behaviour the model exhibits without being shown examples.Thousands of curated, reviewed input–output pairs, plus an evaluation set held back from training.A re-tuning cycle every time the desired behaviour or domain shifts — the heaviest of the three.Deep style, format or domain behaviour retrieval cannot carry.

The honest default

Why most fine-tuning requests end as retrieval projects

The request usually arrives as "we want a model trained on our data", and behind it is a reasonable wish: answers grounded in the organisation's own knowledge. But tuning is the wrong tool for knowledge. A fine-tuned model does not memorise your documents reliably, cannot cite which one an answer came from, cannot respect per-user permissions, and freezes on the day training ends — while your policies and prices keep moving. Retrieval does all four naturally, because the documents stay outside the model and are fetched fresh per question.

Where tuning earns its cost is behaviour, not knowledge: a house style no prompt quite pins down, a narrow classification task where latency and cost push toward a small specialised model, terminology and phrasing in a domain the base model handles clumsily. Those are real cases and we build them. The discipline is that the case is proven by an evaluation first, because a tuned model is an asset you must maintain, and an unnecessary one is maintenance with no return.

The method

How a fine-tuning engagement runs when it is justified

Evaluation opens and closes the work. Nothing is claimed about the tuned model that the held-back test set has not shown.

STAGE 01

Baseline before anything

We build an evaluation set from your real tasks and score the base model — with good prompting and, where relevant, retrieval — so the tuning decision is made against evidence, not intuition. A measured baseline and a go or no-go on tuning.

STAGE 02

Curate the training pairs

Tuning data is thousands of reviewed input–output examples of the behaviour you want, deduplicated and cleaned, with a slice held back that the model never sees during training. A training set and an untouched test set.

STAGE 03

Tune, then re-run the same evaluation

The tuned model is scored on the identical held-back set, like for like against the baseline. If the gain does not justify the ownership cost, we say so and keep the retrieval architecture. A before-and-after comparison you can defend.

STAGE 04

Host, monitor, plan the re-tune

The tuned model is deployed where your governance requires — your tenancy or a governed endpoint — with production sampling reviewed against the evaluation set, so drift is detected and the re-tuning cadence is a schedule, not a surprise. A served model with a drift watch and a re-tuning plan.

Our role

Where we stand in the fine-tune-or-not decision

  1. 01

    An honest recommendation first

    We are paid for the outcome, not the technique, so if the baseline shows retrieval and prompting solve your problem, the recommendation is to stop there — most engagements that start with this question end without a tuned model.

  2. 02

    The tuning done properly when it wins

    Where the evidence supports tuning, we run the data curation, the training and the like-for-like evaluation, and we document what changed so the result is explainable to a reviewer.

  3. 03

    An operable model, not an orphan

    The tuned model ships with its hosting, its monitoring, its evaluation set and its re-tuning runbook, so your team can keep it honest after we leave rather than inheriting a frozen artefact.

What behaviour the model must exhibit, and what quality bar justifies the ownership cost, are your calls; we put numbers from your own tasks behind each option so the call is made on evidence.

Common questions

What buyers ask

When is fine-tuning an LLM actually worth it over retrieval?+

When the thing you need to change is behaviour rather than knowledge, and an evaluation proves the cheaper levers cannot get there. Concretely: a strict house style or output format the model must produce without examples, a narrow high-volume task where a small tuned model beats a large general one on latency and cost, or domain phrasing the base model consistently mangles. If the need is "answer from our documents", that is a retrieval problem, and tuning will do it worse — no citations, no permissions, and knowledge frozen at training time.

How much data does fine-tuning need, and does ours qualify?+

The honest unit is quality examples, not gigabytes. Tuning needs curated input–output pairs that demonstrate the target behaviour — typically thousands of them, reviewed, deduplicated and consistent, with a held-back slice for testing. Most organisations discover their raw material is not yet in that shape: tickets without approved answers, documents without the decision they led to. Part of the engagement is deciding whether usable pairs can be curated from what you have at a reasonable cost, and that finding sometimes settles the tune-or-not question by itself.

Who hosts a fine-tuned model, and how often does it need re-tuning?+

Hosting follows your data governance: a tuned open-weight model can run in your own tenancy or region, while a tuned commercial model lives on that provider's governed endpoint — the right answer depends on what the model will see, and we design it with your compliance owners. Re-tuning has no universal interval; it is triggered by drift. We monitor sampled production outputs against the evaluation set, and when the scores slip — because your products, policies or language moved — that is the signal, so the cadence follows how fast your domain changes rather than a calendar we invented.

Next step

Evaluate it on your data

An evaluation on your own tasks that says, with evidence, whether fine-tuning is worth its cost or retrieval and prompting already get you there.