SIDERIUSSIDERIUS

Tutorial 3 of 3

Monitor runs and read results

Follow a run by its operation ID and tell valid results from progress.

The idea to keep in mind

Finished is not the same as successful, and a score is not the same as a valid result. SIDERIUS keeps progress, scores and validity separate on purpose. A failed or invalid model is still a result worth reading.

Before you start

  • A run you launched, either from the browser console (tutorial 1) or from your project's launch script.
  • For console runs: the operation ID shown at launch.

Source of truth: the candidate console README (opens in a new tab)

Steps

  1. Step 1: Open the run in Slow control

    In the console, choose Slow control and enter the complete operation ID. You can also bookmark a link of the form http://127.0.0.1:8740/#slow-control/<operation-id> and Connect. You do not need to select a task or project to watch a run.

    Closing the page, tunnel or server does not cancel the run. A disconnected view is stale, not stopped.

    Expected result: the run's status, with sections for Operation facts, the Reported token table, and Activity and logs that you can expand.

  2. Step 2: Read the token table for what it is

    These are the providers' own reports. They help you follow spending, but they are not a complete bill, and they say nothing about whether a model is scientifically valid.

    Expected result: usage grouped by requested and served model. Missing values show as Unavailable.

  3. Step 3: Know what each outcome means

    OutcomeWhat it means
    Refusal before startingA required input is missing, invalid or inconsistent. Fix the input; this is not a model result.
    Execution failureTraining, inference, scoring or saving did not complete.
    Health invalidationEverything ran, but the task's own checks declared the output scientifically invalid. Its score is not a best score.
    No valid record yetNo scientifically valid result exists for that iteration. The loop may still continue with that feedback.
    CompletedAn authoritative result was saved for that iteration.

    The full table, with exact status names, is in operating a run (opens in a new tab).

    Expected result: you can place every iteration in one of the rows below.

  4. Step 4: Look inside the workspace

    Every run writes to its own workspace in your project. For command-line runs you can also browse results with the dashboard, from the SIDERIUS checkout:

    .venv/bin/python src/dashboard/main.py     # http://localhost:8000

    Expected result: per-node output records, the effective Health configuration, an invariants lock and the generated model plugins.

After reading the results

To try again with different settings, start a fresh run: a new workspace for a new identity. Resuming keeps the previous run's recorded inputs. Changing the metric or split changes the task; changing iterations or budgets changes only the experiment. Keep both explicit so results stay comparable.