Lab and training
First adapter training run
The adapter walkthrough starts with a sample base model, training dataset and bounded recipe, then follows the resulting candidate record. SAVRN Cloud uses synthetic metrics here to demonstrate lineage and review; no optimization runs and no actual model weights are produced.
SAVRN Cloud · Coming soon · Public preview · Browser-local simulation · No live compute or payments
Complete the local workflow
Open Training and choose New training run. Select a model, a dataset marked Training examples, LoRA or QLoRA, 1–5 epochs, learning rate 0.0001/0.0002/0.0003 and rank 8/16/32. Start the simulation, inspect its state and sample metrics, then follow the generated candidate in Candidates. These are demo choices, not qualified production limits.
This does not execute optimization or create actual model weights. Any loss curve, duration, usage or checkpoint shown is synthetic.
Example recipe boundary
The following is explanatory data, not a released job schema:
{
"simulation": true,
"schema_status": "illustrative",
"base_model": "Research 8B",
"method": "proposed supervised adapter recipe",
"dataset_revision": "synthetic-demo-1",
"production_checkpoint": null
}
Review whether the local result preserves the model, data and recipe context. A candidate record must remain distinct from an approved deployment.
Before connected services launch
Qualify one exact recipe, including tokenizer, adapter method, optimizer settings, sequence lengths and hardware. Establish an enforced budget and resource lifetime. Preserve real metrics, checkpoint digests, code revisions and resource-time accounting. Evaluate against a held-out baseline before promotion. Confirm serving compatibility and rollback. Do not advertise arbitrary training parameters or model sizes merely because the local form demonstrates them; supported values must come from measured acceptance on actual capacity.
Build toward the work that matters.
Tell SAVRN what your institution needs to run, who reviews the results and where its data must stay. That workload defines the next service to qualify.
Discuss an AI project