Admin Settings → Training connects and governs the fine-tuning capability users access from the Fine-Tuning module.
Enable the integration
| Step | Action |
|---|---|
| 1 | Enable the training integration. When disabled, users see that training is unavailable. |
| 2 | Base URL — the address of the fine-tuning service in your environment. |
| 3 | Bearer token — the optional credential the service requires. |
| 4 | Connection config (JSON) — advanced connection configuration where required; it must be valid before saving. |
| 5 | Verify connection — tests reachability before you save. A green verification confirms the fine-tuning engine is ready. |
Note:
Fine-tuning is GPU-intensive. Schedule large jobs via your change process, and watch GPU Monitoring during runs so training doesn't starve interactive chat of capacity.
Who may train
Training access is a permission like any other — grant it to the groups who need it via Permissions. Most organisations limit it to an ML/platform team initially.
Governance tips
- Data — training data stays in your environment; still, treat datasets as sensitive and limit who can create them.
- Lifecycle — retired tuned models should be removed to free disk (Cluster Manager → Storage).
- Quality gate — require a scored Comparison or evaluation before a tuned model enters the default list. See Models (Admin).
Troubleshooting
| Message | Meaning |
|---|---|
| "LlamaFactory URL is required before verification" | Enter the service address first. |
| "Config JSON must be valid JSON before saving" | Fix the advanced configuration syntax. |
| "Failed to verify … connection" | Service down, URL wrong, or token invalid. |