Fine-tuning consultancy or provider tools: when to hire help

The training run is the cheapest and most automated part of most fine-tuning projects. Managed platforms take a file of examples, a base model and a few settings, and return a tuned model. What they do not do is decide whether those examples are the right ones, or whether the result is better than what you had. That gap is where a consultancy earns its fee, or fails to.

What provider tools actually cover

A managed fine-tuning platform (a model provider’s API, a cloud ML service or an inference company’s training product) typically handles:

• uploading and validating a dataset in its required format;
• running supervised, preference or reinforcement tuning on a menu of base models;
• reporting training and validation loss;
• serving the tuned model on an endpoint, or letting you download weights where the license allows.

The price is published and usually low for small jobs. Some platforms bill per million training tokens, some per GPU hour, and a few set a minimum per job.

What they leave to you

• Choosing and cleaning examples. Duplicates, contradictions and test items copied into the training set quietly ruin a run.
• Building an evaluation. A held-out test set, a rubric and a baseline to beat.
• Writing a grader for reinforcement fine-tuning, and checking it rewards the right thing.
• Deciding whether to tune at all. Prompting or retrieval may solve the problem for less.
• Serving economics. Matching traffic to endpoint type and hardware.

If your team can do those five things, a consultancy adds little beyond extra hands.

Signals that provider tools are enough

Situation	Why self-serve works
You have a few thousand reviewed examples in a consistent format	The data work is already done
You have a test set and a number to beat	You can tell if the run helped
The task is narrow: format, tone, classification, extraction	SFT on a small model often handles it
Hosting on the provider’s endpoint is acceptable	No deployment project on top
An engineer has a week to iterate	Small runs can be repeated cheaply while you tune settings

Signals that you should hire help

Situation	What a consultancy brings
You have raw logs or documents, not labeled pairs	Data design, labeling guidance, cleaning
Nobody agrees on what “good” looks like	A rubric and an evaluation set built with your domain experts
You want reinforcement fine-tuning	Grader design and reward-hacking checks
The model must run in your VPC or on premises	Serving, quantization and infrastructure work
A security or procurement review needs documentation	Written method, data flow and test results
You have tried once and it did not beat prompting	A diagnosis of why

The middle option

Many teams end up mixed: a consultant designs the data and the evaluation, then the team runs training on a provider platform it already pays for. This keeps the recurring cost (serving) on a published price list and limits consulting to the part that needs judgment. It also leaves you able to retrain without calling anyone back.

If you go this way, put the evaluation set and grader in the deliverables, owned by you. They are worth more than any single tuned model, because every future run is measured against them.

How consultancies price

Expect one of three shapes:

1. Fixed-fee discovery (often a few weeks): assess the data, build a baseline, recommend whether to tune.
2. Project fee against a statement of work: data, training, evaluation, handover.
3. Retainer or day rate for ongoing retraining and monitoring.

Few consultancies publish rates, so compare on scope instead: what is delivered, who owns it and what counts as done. Our post on what a fine-tuning statement of work should include lists the clauses to look for.

Questions to settle before choosing

• Do we have examples, or do we have data that could become examples?
• Can we say in one sentence what the tuned model should do better, and how we would measure it?
• Where must the model run, and who will keep it running?
• Is this a one-off, or a model we will retrain every quarter?

If the honest answers are “examples, yes, on the provider, one-off”, start with provider tools. If any answer is “not sure”, that uncertainty is the project, and it is worth paying someone who has resolved it before.

The vendor notes cover what each platform on our list publishes about methods, evaluation and deployment, which helps with the first half of this choice. For the second half, our list of the best LLM fine-tuning companies separates platforms from firms that do the work for you.

Pick the option that leaves you owning the test set.