AI Services
LLM Fine-Tuning
A legitimate technique that is reached for far too early. Most problems people bring to fine-tuning are prompt or retrieval problems.
Fine-tuning adapts a commercial model to your data and format. The Nexclick recommends it only after prompting and retrieval have genuinely been exhausted, because those solve most problems at a fraction of the cost and without the maintenance a tuned model carries.
Is this you?
What usually prompts the call
- You need consistent output in a specific format and prompting gets it right most of the time.
- Your domain uses terminology and conventions general models handle poorly.
- Prompt length has grown so large that cost per call has become significant.
- A vendor has proposed fine-tuning and you cannot tell whether it is justified.
What we do
The actual deliverables
Things that appear on an invoice, not adjectives.
- Exhaust the cheaper options first
- Better prompting, few-shot examples and retrieval solve the overwhelming majority of problems people bring to fine-tuning, at a fraction of the cost and with no maintenance burden.
- Assess whether you have the data
- Fine-tuning needs consistent, high-quality examples — usually hundreds at minimum. Tuning on inconsistent examples teaches the model to be inconsistent.
- Build an evaluation set first
- How you will judge whether the tuned model is better, agreed before tuning. Without it the comparison becomes an impression, and impressions favour whatever was built most recently.
- Prepare the training data
- Cleaning, formatting and deduplicating examples. This is most of the work and most of the cost, and it is what determines the result.
- Tune and evaluate honestly
- Against the baseline, on the evaluation set. If the tuned model is not measurably better, we will say so — that outcome happens and it should be reported.
- Cost comparison at your volume
- Tuned models can be cheaper per call with shorter prompts, or more expensive depending on the provider and tier. Modelled at your actual usage rather than assumed.
- A maintenance plan
- Tuned models are pinned to a base version. When providers deprecate that version you retune, and that cost belongs in the decision from the start.
Decision tree
Do you actually need fine-tuning?
Fine-tuning is reached for early and justified late. Work down in order — the first four branches all point at something cheaper that solves the problem more reliably.
01You want the model to know your company information
Retrieval, not fine-tuning. Tuning does not reliably teach facts and it cannot be updated when the information changes. This is the most common misapplication by a distance.
02The output is nearly right but occasionally wanders off format
Better prompting with few-shot examples, and structured output constraints. Usually fixes it in an afternoon.
03It handles common cases well and fails on edge cases
Add those edge cases as examples in the prompt. Cheaper, immediate, and adjustable without a retraining cycle.
04Your prompts have grown very long and cost per call is climbing
A legitimate reason. Tuning can encode what the long prompt was teaching, reducing tokens per call. Model the cost saving before committing.
05You need a consistent tone or format across thousands of outputs
Good candidate. Tuning is genuinely effective at format and style consistency, which is what it does best.
06Your domain has terminology and conventions general models handle poorly
Reasonable candidate, if you have several hundred consistent examples. Without them, tuning will teach the inconsistency.
07A vendor recommended it and could not explain what prompting failed to achieve
Ask that question directly. If prompting and retrieval were not properly attempted first, the recommendation is premature.
How it works
Step by step, with timeframes
Timeframes are typical rather than guaranteed, and they assume we get account access and approvals when we ask.
- 01Week 1–2
Try the cheaper options
Prompt engineering, few-shot examples and retrieval, measured against your requirement. This frequently ends the engagement, which is the correct outcome.
- 02Week 2–3
Build the evaluation set
Test cases with known-good outputs, agreed before tuning, so the comparison is objective.
- 03Week 3–6
Prepare data and tune
Training examples cleaned and formatted, then tuning runs with results measured against the baseline.
- 04Week 6–8
Deploy or report the negative
Deployed if measurably better, with monitoring. Reported honestly if not — a negative result is a valid deliverable.
What you get
Reporting and ownership
- A documented attempt at the cheaper options first, with results, before tuning is recommended.
- An evaluation set agreed before tuning, so the comparison is objective rather than impressionistic.
- The prepared training dataset, owned by you and reusable with any provider.
- A measured comparison against the untuned baseline, reported honestly either way.
- A retuning plan for when the base model version is deprecated.
Tools and platforms
- OpenAI and Anthropic fine-tuning APIs
- Training data preparation pipelines
- Evaluation frameworks
- Baseline prompting and retrieval comparison
- Cost modelling at volume
Timeline
How long this actually takes
Six to eight weeks, and the first two frequently end it — prompting and retrieval solve most problems brought to fine-tuning. Two honest positions. Fine-tuning teaches format, tone and consistency; it does not reliably teach facts, and using it to make a model "know" your information is the most common misapplication. And a tuned model is pinned to a base version that providers eventually deprecate, so retuning is a recurring cost that belongs in the decision rather than arriving as a surprise in eighteen months.
Pricing model
Fixed-price project
Fixed price including the cheaper-options assessment, which stands alone. Tuning quoted separately and only recommended where the assessment justifies it.
Questions
LLM Fine-Tuning questions
Will fine-tuning make the model know our information?
Not reliably, and this is the most common misunderstanding. Fine-tuning teaches format, tone and behaviour. It does not implant facts dependably, and it cannot be updated when the facts change. For knowledge, use retrieval — that is precisely what RAG exists for.
How much training data do we need?
Usually several hundred consistent, high-quality examples at minimum, and consistency matters more than volume. Tuning on inconsistent examples teaches the model to be inconsistent. If you cannot produce that dataset, the assessment stage will say so before any cost is committed.
Is a fine-tuned model cheaper to run?
It can be, where it lets you use much shorter prompts. It can also be more expensive per token depending on provider and tier. We model it at your actual call volume — the answer varies enough that assuming either direction is unwise.
What happens when the base model is deprecated?
You retune on the new version. Providers do deprecate base models, and a tuned model is pinned to the one it was built on. That recurring cost belongs in the original decision rather than arriving as a surprise a year later.
What if the tuned model is not better?
We report it. That outcome happens, which is exactly why the evaluation set is agreed before tuning rather than after. A negative result measured properly is a legitimate deliverable and considerably cheaper than deploying something on an impression.
Can we fine-tune an open-source model instead?
Technically yes, and it brings hosting, scaling and maintenance obligations most businesses underestimate. For the majority of commercial use cases, fine-tuning a hosted commercial model is simpler and cheaper in total cost. Self-hosting earns its place on data residency or volume grounds.
Last reviewed 28 July 2026.
Tell us what you are trying to fix
A 20-minute call, no pitch deck. The Nexclick will tell you what we would do, roughly what it costs, and whether we are the right people for it.