Fine-tuning trains a model on your own examples so it responds in your format, your voice, or your domain — without hand-writing that guidance into every prompt.

Upload training data

Under AI Gateway → Fine-tuning, upload a CSV with a header row and two columns: prompt and completion. Each row is one training example — the input you’d send, and the output you want back.

The upload is validated before training starts; a missing header or missing column is rejected there, not partway through the job.

Launch and deploy

Review the quote and choose Accept and start training from the same screen. Training runs in the background — track its status (uploading, training, deploying, ready, or failed) in the dashboard.

Once a job reaches ready, the custom model deploys automatically under a custom-* model id. From there it behaves like any other model in your catalog:

  • It appears in the playground model picker.
  • It’s callable through GET /v1/models and POST /v1/chat/completions, same as a stock model.

If a status check is interrupted, the job stays visible and SetCloud retries it automatically. Wait for an active job to finish before deleting the model.

A failed job reports its error in the dashboard — check the failure message before re-uploading.

How it bills

Choose a base model and validate your dataset, then review and accept the fixed job quote before starting. The quote is based on the UTF-8 size of the normalized training dataset (1 MB = 1,000,000 bytes), covers two passes, and rounds up to cents. It is charged once after successful training. Add enough credits to cover the quote.

Custom model storage is billed from successful training until deletion, prorated by time. Inference is billed separately per request at the model rate. Current training, storage and inference rates are on Pricing. Deletion removes the finished model and its inference deployment.

More: