Dashboard Training
Use the Halo Forge dashboard as the primary operator surface
The dashboard is the normal interactive path for local workstation training. The CLI remains an equal automation path over the same catalog and services. Both support choosing a goal, checking preflight, launching, monitoring, evaluating, serving, and inspecting artifacts without changing data formats.
On ROCm and CUDA workstations, Setup first prepares and qualifies a pinned managed runtime. Train then uses that runtime automatically. A detected GPU is never shown as ready until the runtime has completed a real optimizer update and saved-artifact reload. See Managed Accelerator Runtimes.
Guided vs Advanced Train
- Train → Guided is the beginner-safe path with conservative defaults and structured own-data selection.
- Train → Advanced exposes direct method configuration for SFT, RAFT, DPO, ORPO, RM, GRPO, VLM, audio, reasoning, and agentic training.
- Runs monitors active and completed work and owns completed-run actions.
- Models owns trained artifacts, serving, conversion, qualification, and export.
Default Output Path
Dashboard launches save under:
~/.halo-forge/runs/<method>-<goal-or-template>-<model-slug>
This avoids installed-app permission failures from repo-relative models/... paths.
Method Preconditions
| Method | Needs |
|---|---|
| SFT | model, dataset, writable output path |
| RAFT | model, prompt file, verifier |
| DPO/ORPO | model, preference dataset |
| RM | model, preference dataset |
| GRPO | model, prompt dataset, verifier |
| VLM | compatible VLM family and image-text data |
| Audio | audio dependencies, task, audio data |
| Reasoning | compatible text model and reasoning data |
| Agentic | tool-call traces or structured-output data |
When a method is capability-gated, the dashboard shows the reason and keeps the CLI path documented. The guided picker only exposes methods certified for the active runtime, trainer adapter, and capacity adapter.
Serving After Training
When a run produces a final model or adapter, open Runs → Artifacts or the shared Models library and choose Serve. Halo Forge manages one local serve process at a time and sends Models → Serve & Test to the managed endpoint by default.