Skip to main content

Baselines

Open Assessments > Drift & Bias > Baselines. The tab has two sections.

01 Text models: LLM baselines​

A baseline is the first complete run of a series: the same model, question set, settings and judge. Later runs of the series are compared with it. Baselines are kept until deleted; the next complete run of the series then becomes its baseline. A run marked Incomplete (more than 10% of its questions could not be sent or judged) is never used as a baseline.

02 Classic models: ML baselines​

Figure 1. The Baselines tab.

Each model version is compared with its reference window:

  • The rows your pipeline sent as reference (shown as Sent as reference), or
  • Its first 7 days of scored rows, when no reference rows were sent.

A new version gets its own reference.

ColumnShows
ModelName and model key.
VersionThe version this reference belongs to.
ReferenceWhere the reference came from.
WindowThe dates the reference covers.
RowsRows in the reference, and how many carried a label.
Last checkedThe model's latest check.

Both sections have an Export button that opens a new export in Export Center. See Export.