Chapter 5 — Dashboard Overview
Bench — the live dashboard
Start the Bench on a run directory with iris /path/to/run-dir. It reads
the diagnostic bundle while training continues or after training has ended.
You can inspect the same captured run without loading its model again.
For launch flags and remote access, see Chapter 10.
Find the view that answers your question
| View | What to inspect |
|---|---|
| Run | Run identity, captured frames, credits, resource readings, and tracker controls. |
| Concept Space | A projection of how labeled probes are organized in the recorded representations. |
| Whole Model | Training outcomes and internal structure or motion across the run. |
| Domains | Differences between the categories assigned to your probes. |
| Layers | Spectra, effective dimensionality, orientation, and other layer-level readings. |
| Heads | Per-head readings when the architecture and capture support them. |
| Experts | Expert and routing readings on supported mixture-of-experts captures. |
| Facets | Observed groupings of probes, with explicit evidence and identity limits. |
| Flow Map | Layer-to-layer signal views, where the run recorded them. |
| Run Notes | Your record of interventions and observations. |
| User | Signals you reported and functions you define over available channels. |
| AI Review Prompts | Prompts for integration, probes, and evidence-aware report reading. |
| Manuals | Product documentation inside the Bench. |
Not every model or bundle supports every view. Read the sentence explaining an unavailable panel before changing settings. An architecturally inapplicable measurement differs from one that was not captured, was shed, or is still being prepared.
Compare the same frame
The playback bar lets you inspect recorded training steps across views. Pause when reading exact values. A display animation does not create new measured frames between captures. Look at step labels and capture gaps, especially when comparing an internal change with loss or accuracy.
Keep observations with the run
Use Run Notes and chart observations to record what you saw and what changed in the training setup. Note checkpoint resumes, optimizer changes, data changes, and manual interventions. A timestamped observation helps you evaluate a proposed explanation later; it is not itself evidence of cause.
Your own signals and functions
Report a scalar with tracker.report_signal(index, value, label=..., unit=..., description=...). In the User tab, Reported signals remain
distinguishable from IRIS™-derived measurements. Give each signal a clear
definition and unit so another reader can interpret it.
Derived functions are evaluated in the Bench over data already in the run. Use the variable picker to choose available channels and save expressions you want to reuse. A function cannot manufacture an absent channel. If inputs have no jointly available samples, preserve that gap rather than interpolating a result and treating it as measured.
Check cost and resources
Review the run's license and credit state before training, and use the portal for current pricing and account requirements. Measure instrumentation overhead on your own model and hardware. Capture cadence, probe count, available diagnostics, and resource pressure affect the work performed; neither another run's overhead nor a resource badge predicts yours.
