Your data and fine-tuning

How it works

1. Upload a dataset to a Tenant. rackAI accepts JSON Lines in chat, ShareGPT, Alpaca or completion format, up to 256 MiB per file. Every line is checked as it arrives. If a line is wrong you'll see which one, and nothing from a rejected upload is kept.
2. Start a fine-tuning job with LoRA or QLoRA on a model from your catalogue, on 1 to 8 GPUs. Settings you leave unset use sensible defaults, and the job reports every value it used.
3. Get an adapter. A finished job produces a LoRA adapter: a small set of weights that sits on top of the base model.
4. Serve it. Add the adapter to a deployment of the same base model. Several adapters can be served beside their base model at once. Call one by name in the model field of your request.

What happens to your training data

• It stays in your Tenant. A dataset belongs to the Tenant that uploaded it. Only that Tenant's fine-tuning jobs can read it.
• It doesn't change once uploaded. Datasets are immutable, so a job always trains on exactly what you uploaded. You can download your file again at any time.
• Training runs in a closed room. Jobs run on GPU capacity assigned to your Organization, with no network access. They can read only that job's dataset and base model.
• It isn't used to train anyone else's model.

Adapters and base weights

• Base model weights are never modified. They are mounted read-only. Fine-tuning writes only the new adapter.
• Adapters stay in the Tenant that trained them. A deployment can serve only adapters trained in its own Tenant, on the same base model.
• Adapters in use are protected. You can't delete an adapter while a live deployment is serving it.

Retention and deletion

• You decide when training data goes. Datasets, fine-tuning jobs and their adapters are kept until you delete them.
• Deletion removes the files. Deleting a dataset removes the stored file and its record. Deleting a job stops it if it's still running and removes its adapter and logs. Each deletion is recorded in your activity trail.
• Safe by default. A dataset can't be deleted while a job is still training on it.
• Deleting a Tenant removes its datasets, adapters, models and deployments.
• Usage and activity records are kept for the retention period your Organization sets, then purged automatically.

What fine-tuning costs

During the trial, fine-tuning jobs don't use your trial tokens. Pricing for training is not yet published.