| name | trackio-observability |
| description | Use when instrumenting or inspecting TRL training runs with Trackio, run names, metric schemas, dashboards, logs, grep or ripgrep, SFTP, Hugging Face Job logs, remote artifacts, or experiment result summaries. |
Trackio Observability
Use this skill to make training runs observable and debuggable.
Workflow
- Define the run identity: project, run name, method, model, dataset, seed, and
challenge.
- Add Trackio. For any remote Hugging Face Job, initialize a hosted dashboard
with
trackio.init(..., space_id="owner/space"); only short local smoke
tests may stay local or skip tracking with an explicit reason.
- Make logs grep-friendly with clear phase markers.
- Persist artifacts intentionally: model, adapter, config, metrics, traces, and
evaluation outputs.
- Inspect remote state with the narrowest tool: Trackio dashboard, HF CLI,
rg, or SFTP when configured.
Reporting Shape
Return:
- run id or job id
- Trackio dashboard or Space
- command or script inspected
- latest metrics
- artifact paths
- failure signatures
- next minimal action
Never print tokens, secrets, private credentials, or full logs unless the user
explicitly asks for a raw excerpt.
References
references/tracking-schema.md: run metadata and metric schema.references/log-inspection.md: grep, SFTP, and remote artifact triage.
