# Lab and learning methods

Canonical: https://savrn.com/cloud/docs/lab-overview

SAVRN Cloud · Coming soon. Public documentation and a browser simulation are available; connected services are not yet available.

Model improvement can start with clearer instructions, better retrieval or a corrected evaluation before training is justified. SAVRN Cloud's lab preview connects these choices to candidate review, making the distinct requirements of supervised tuning, adapters and reinforcement learning explicit.

**SAVRN Cloud · Coming soon · Public preview · Browser-local simulation · No live compute or payments**

## Follow the improvement loop

Open [Training](https://savrn.com/cloud/console/#/training) to inspect a demonstration run and its configuration. Follow its output into [Candidates](https://savrn.com/cloud/console/#/candidates), review related [Evaluations](https://savrn.com/cloud/console/#/evaluations), then inspect the promotion path into [Deployments](https://savrn.com/cloud/console/#/deployments). Every metric, checkpoint description and generated candidate is synthetic.

The local walkthrough makes the release process visible. It does not change model weights, allocate training hardware or establish a quality improvement.

## Choose the least complex useful change

A task may improve through prompt clarity, retrieval, tool design or an evaluation fix before training is needed. Supervised tuning uses approved examples; LoRA is an adaptation technique that can support such training. Reinforcement learning needs a validated reward process. Full parameter tuning and distributed training impose separate capacity and operational requirements.

These are distinct methods, not interchangeable checkboxes. A model fitting in memory for inference does not prove that its training workload fits.

## Before connected services launch

Begin with one qualified recipe and a small approved dataset. Freeze the base model, tokenizer, training code, settings and splits. Capture actual metrics and checkpoint identity, then compare against a held-out baseline. Promotion requires an authorized decision and a controller that observes the new serving identity. Retain rollback. Public support for one recipe should not imply arbitrary-model training or all learning methods.