Choose an approved base model. Noire locks the exact source and commit SHA so every training and validation result remains reproducible.

Validation stores intentional test inputs and a cryptographic output digest, never raw generated output.

The worker generates actual completions from the selected immutable revision and stores the scorecard.

The prompt is sent to the active revision and is not retained by Noire.

Waiting for a prompt.

Record a reviewable model change. Training starts only after you select a method.

The commit author is recorded from your signed-in account.

LoRA trains a compact, reversible adapter on the model change. It is the fastest and lowest-risk default for iterative knowledge updates.

Noire creates a seven-day invitation link. Add an email to reserve it for one person, or leave it empty to create one shareable link for this organization.