Fine-tuning
When prompting and retrieval aren't enough — the model needs your domain's vocabulary, your format, your judgment — fine-tuning trains the model itself on your examples. Pradra Studio runs it as a governed job with the same discipline as everything else: consent, budgets, and the same quality gate before anything serves.
Two paths, one lifecycle
| Path | Where training happens | Where your data goes |
|---|---|---|
| Self-hosted (LoRA/QLoRA) | Inside the platform | Nowhere — weights and data stay home. |
| Provider API | The model provider's service | Uploaded to the provider — hence the consent gate below. |
Both paths submit the same way, are tracked the same way, and produce a candidate model that faces the same gate.
The consent gate
The provider path uploads training data to a third party. That is refused — by default, for every provider — until an organization administrator explicitly grants consent for that provider. The grant is recorded like a license acceptance (who, when, which provider) and can be revoked at any time; revoking stops new jobs without affecting finished models.
Cost is booked honestly
Fine-tuning draws from the same monthly budget as generative inference. A job's cost is booked even if the job fails — because the provider charged you either way, and a budget that ignores failed spend isn't a budget.
No new promotion path
A finished fine-tune does not ship itself. The result becomes a candidate model that must be pinned to the capability and pass the same golden-set evaluation and acceptance criteria as any prompt change. A fine-tune that fails its golden set cannot reach production, and rolling back is the usual pointer switch to the previous model.
Working with fine-tuning in the Studio
- Go to Fine-tuning in the sidebar.
- The consents card shows which providers this organization has approved — grant or revoke there (admin only).
- Submit a job: pick the capability, the path (self-hosted or provider), the base model, the training dataset version, and hyperparameters.
- The jobs table tracks each run to a terminal state. A finished job records the resulting model; from the capability's Model tab, pin it, evaluate it, and activate it through the gate.
When to fine-tune (and when not to)
- Reach for prompt work first. A better prompt version costs minutes and no training; see Prompts.
- Reach for retrieval second. If the problem is missing knowledge, a knowledge source fixes it without touching the model.
- Fine-tune for behavior, not facts — tone, format discipline, domain shorthand — and only once you have hundreds of good examples, because the golden set that gates the result needs to be trustworthy too.
Next
- The gate every candidate faces: Generative evaluation.