jolibrain Training Observability
Use training-observe against only the run root the user has placed in scope.
The command reads the run; render writes solely to its explicit output path.
Inspect a run
Run
training-observe summary RUNto establish status, latest scalar values, plot IDs, visual-artifact availability, and reader warnings.Run
training-observe metrics RUN --name METRICfor the raw points that support any claim about a trend. Treat elapsed-time and progress fields as operational signals, not model-quality metrics.Run
training-observe plots RUN, then render only relevant plots, for example:training-observe render RUN --plot loss-train-loss --output /tmp/train-loss.pngInspect the emitted PNG with the available image-viewing capability.
Run
training-observe artifacts RUN --step latestand inspect both each returned image and its paired JSON metadata before judging prediction quality.
Report evidence
- State the exact metric names, step range, and values that support a trend.
- Separate observations from inferences. Report a loss plateau before inferring that an optimizer or data change could help.
- Compare per-test-set traces independently; do not average them by eye.
- Treat missing plots or results as missing evidence, not a good or bad result.
Preserve control boundaries
Do not start, stop, cancel, resume, delete, or reconfigure a training run. Do not scrape the live Visdom UI as the source of truth. Use the flushed metric stream and saved visual artifacts, which remain available while the run is live and after the Visdom server exits.