Training workspace

Loading

Select a run to explore its progress and results.

Elapsed—
Fit ETA —
Macro F1 —

Readiness for review

Waiting

Research comparisons, not an automatic release decision. Acceptance thresholds have not been agreed; prospective validation is still required.

Held-out performance

Folds

One night group held out at a time

Per-stage F1

Held-out, pooled

Candidate selection is repeated inside each fold. The candidate can differ between nights.

Confusion vs Watch

Watch ↓ · Prediction → · Row-normalized. Watch is a reference, not clinical ground truth.

Feature importance

Run health

Last 80 fit durations. Measures work, not model quality. Memory is the trainer’s recorded peak, not current system usage.

Config

Run notes

All model comparisons

Individual candidate scores are exploratory. Rules may have a different number of matched labels. Always light is a simple reference that predicts light sleep every time.

Learn the terms

Actions

Jobs