Questions to ask a fine-tuning vendor before you sign
Published by Best LLM Fine-Tuning Companies
Most fine-tuning proposals look alike on the first page. The differences sit in units, defaults and contract clauses that nobody reads until an invoice or a security review forces the issue. These are the questions we would put to any vendor, grouped by the part of the project they protect.
Price: what exactly is being counted?
Fine-tuning is priced in at least three different ways, and the same project can look cheap or expensive depending on which one you are reading.
| Price basis | What it counts | What to ask |
|---|---|---|
| Per training token | Tokens in the dataset, times the number of epochs | How many epochs are assumed, and are validation tokens billed too? |
| Per GPU or training hour | Wall-clock time on named hardware | Which GPU, how many, and who pays for a job that fails? |
| Fixed fee or day rate | A consultancy’s time | What is in scope, and what triggers a change order? |
On token pricing, the unit itself varies. Google’s Vertex AI pricing page lists Gemini tuning per 1,000 training tokens while open models on the same page are priced per million. Ask the vendor to restate its estimate per million training tokens so you can line quotes up.
Then ask about minimums. Some platforms set a minimum charge per job, which matters if you plan many small experiments rather than one large run.
Models: which base models, and for how long?
A tuned model is tied to its base model. If the base model is retired, the tuned version can go with it.
- Which base models can be tuned today, at which sizes?
- What is the published deprecation policy for those base models?
- Can you export the tuned weights, or only call them through the vendor’s API?
- If the base model is open-weight, which license applies to the tuned result?
The answer to the export question shapes everything after it. A model you can download can move to another host. A model that only exists behind one provider’s endpoint cannot.
Method: what kind of tuning is on offer?
Supervised fine-tuning (SFT) on example pairs is the default. Preference methods such as DPO, and reinforcement fine-tuning against a grader, are now listed by several providers. LoRA adapters are cheaper to train and serve than full fine-tuning, and some providers price the two differently.
Ask which methods the vendor supports for your chosen base model, not in general. A method listed on the marketing page may only be available for two model sizes.
Evaluation: how will we know it worked?
This is the question that separates a vendor from a GPU rental.
- Does the platform run evaluations during or after training, or is that left to you?
- Can you bring your own test set and scoring rubric?
- Will the vendor report results against your current baseline (prompting or retrieval on the untuned model)?
- Who writes the grader if reinforcement fine-tuning is used, and who checks the grader?
A vendor that cannot describe its evaluation step in a paragraph is asking you to trust the loss curve. That is not enough to ship.
Data handling: what happens to the training set?
Get these answers in writing, ideally in the order form or data processing agreement rather than a FAQ page.
- Is training data used to improve the vendor’s own models? Is that the default, or something you opt into?
- How long are uploaded files kept after the job, and can you delete them on demand?
- Where is data stored and processed (region and sub-processors)?
- Who on the vendor’s side can see the data, and under what controls?
Watch for discounts tied to data sharing. A lower inference price in exchange for letting the provider use your data is a trade some teams accept and many security reviews will not.
Deployment: where will the model run, and at what cost?
Training is a one-time cost. Serving is monthly, and it often decides the total.
- Is the tuned model served at the same per-token price as the base model, or at a premium?
- Does it need a dedicated endpoint billed by the hour or by GPU-second?
- Is there a storage or hosting fee per custom model per month?
- Can it run inside your cloud account, or on servers you operate?
Ask for one worked example: your expected monthly traffic, priced on the vendor’s published rates.
People: who does the work?
For a consultancy or managed service, ask who will be on the project, whether they have tuned your chosen model family before, and whether a subcontractor will handle any of it. Ask for a sample deliverable (an evaluation report or a statement of work), not a case study logo.
A short checklist to send
If you only send one email, send this:
- Price per million training tokens (or per hour, with hardware named), minimum charge and epoch assumptions.
- Supported base models, methods and deprecation policy.
- Evaluation tooling and what you will receive at the end.
- Data retention, training-use default and storage region.
- Weight ownership and export.
- Serving price and hosting fees for the tuned model.
We use the same six areas to build our comparison of the best LLM fine-tuning companies, and the scoring method explains how each one is judged from published pages alone.
Read the answers side by side, in the vendor’s own words, before anyone books a kickoff call.
Get a quote
Start with a ZIP code, add an email, then tap through what you want tuned. We read it, check which vendors' published terms fit, and write back with names before anyone else sees your brief.
This is an introduction service. The vendor you are introduced to may pay it a fee if you sign work together; the team asking is never charged.
Brief received.
- We read it against the published terms, model lists and deployment options of the vendors on our scorecard.
- We write back with the vendors that fit and why, by name. Nothing has gone to any of them yet.
- Want an introduction? Tell us which vendor. Your brief and contact details move only after that reply.