ML Issues

What you’ll learn

• How to tell an empty state from a real problem on the ML Center
• Where the placeholder text on each subpage actually comes from
• When to refresh and when to wait

Read this first

The ML Center has many empty states by design. A model with no reviews shows ”—”. A new population with no junk-bucketed detections shows “No rejected detections recorded yet.” A queue with no debounce timers active shows “No fusion fires pending.” None of these are errors — they are accurate steady states.

Before reporting an issue, match what you see against the table below.

Common placeholder states

”—” instead of a percentage

formatMetric renders null or undefined numeric metrics as — (a single em-dash) so you can tell no data from a real 0% or 100%. If a column shows —:

• The model has no reviewed annotations contributing to that metric, or
• The metric isn’t applicable to that model (e.g., a side-classification metric on an identification model)

Wait for reviews to accumulate, or pick a different model — there isn’t an action you take to “fix” this.

”Preliminary” badge

The badge appears next to a model row when the count of reviewed annotations is below the population’s metricSampleThreshold (default 100). The tooltip reads:

Fewer than 100 annotations. Metrics are preliminary.

The threshold is configurable on the population’s ML settings — bigger populations may want a higher bar. Below it, swings of a few reviews change the metrics noticeably; above it, the numbers stabilise.

”Last updated” is in the past

ML Center metrics are materialized — a nightly job aggregates revisions into model-level snapshots. The page reads those snapshots; it does not recompute on load. So:

• Reviews you just did won’t appear until the next nightly run
• The header shows the timestamp the snapshot was last refreshed
• An admin can press Refresh to queue an immediate recomputation; you’ll see the snackbar Metrics refresh queued. Results will update shortly. Watch the timestamp to know when it lands.

If Refresh fails, the snackbar reads mlCenter.refreshFailed — usually a transient backend error; retry.

Fusion shadow: “No shadow data yet for the last N days”

The fusion potential-lift card shows this exact message when no encounter has been scored under shadow mode in the configured window. The remediation is the literal next sentence in the message:

Run an encounter through the pipeline to populate this.

It’s not an error — it just means the data hasn’t accumulated yet on this population.

Junk Review: “No rejected detections recorded yet”

This empty state on the Junk Review page reads:

No rejected detections recorded yet. Buckets fill as gate models reject upstream detections.

A gate-model rejection is what populates the junk buckets. If your population has no gates configured, or gates haven’t fired, this list will stay empty — and that’s expected.

Fusion debounce queue: “No fusion fires pending”

When the debounce queue is empty, you see:

No fusion fires pending. Confirm an annotation to see one queued here.

The queue polls every 2 seconds. Confirming any annotation in the annotator should produce a queued entry within a few seconds — if you’re testing the system, this is how to verify it.

When something genuinely looks wrong

If you see a state that doesn’t match any of the above, capture:

• The model name and id (visible on the row)
• The exact text of any empty state or error
• The Last updated timestamp shown on the page
• Whether refreshing changes anything

For metric definitions (what each percentage actually counts), don’t troubleshoot by guessing — the metric pages document each formula:

• Detection Metrics — precision, recall, IoU
• Classification Metrics — side accuracy
• Identification Metrics — top-1, top-N, mAP@K
• Revision Rate — overall human-correction effort

Related

• ML Center Overview — what the page is and how the metrics are computed
• Common Issues — generic HTTP error toasts (Refresh failures often surface here too)