Drift Radar
A model is trained on the world as it was. Drift Radar answers whether that world has moved — whether the distribution of scores this server produces has shifted away from what it produced when it was healthy.
The baseline comes first
Drift is always relative, so there is nothing to measure until a reference exists.
The baseline is built from this server's first completed batch run.
| Baseline status | Meaning |
|---|---|
| Not started | No batch run has completed yet. |
| In progress | The baseline is being built. |
| Ready | Drift can be measured. |
| Failed | Retry it from this page. |
If the radar is empty, check the baseline status before assuming there is no drift — "no baseline" and "no drift" look nothing alike but are easy to confuse at a glance.
Reading the radar
Each of the model's targets is classified for the selected window:
| Class | Reading |
|---|---|
| Normal | The distribution is where it was. |
| Attention | Moving. Worth watching, not yet acting on. |
| Critical | Materially shifted. |
The summary reports how many of the model's targets are drifting in the window, and you can filter the view down to drifted targets only.
Drilling into one target compares it two ways, and the difference matters:
- vs Reference — against the original baseline. Answers has the world moved since this model was healthy?
- vs Previous — against the preceding window. Answers is it moving right now?
A target can sit far from its reference while being perfectly stable against the previous window: it moved once and settled. That is a different problem from one drifting a little every window, and only the second is accelerating.
The drill-down also exposes the decision cut and alarm counts against the reference.
Acting on drift
Drift is a signal, not a verdict. A shifted score distribution can mean the population changed, an upstream feed changed, or the model has genuinely gone stale.
When retraining is the answer, Trigger Retraining starts the loop from here: pick the base checkpoint to adapt and the labeled dataset to adapt it on. That is an ordinary fine-tuning run — evaluate its result on the leaderboard before promoting it, exactly as you would any other.
Next steps
- Batch runs — where the baseline comes from.
- Fine-tuning · Evaluation