Skip to main content

About Drift & Bias

Drift & Bias tells you whether your models still behave the way they did when you approved them. It covers two kinds of model side by side:

  • Text models (LLMs): how text models from your model library treat gender, race and ethnicity, and age, and whether that changes from run to run. Each run reports a bias-free score, the share of replies without biased opinions.
  • Classic ML models: classifiers and regressions you run in your own pipelines, checked on their own scored rows for drift, performance and fairness.

In the console, open Assessments > Drift & Bias. It sits next to Red Teaming.

The tabs​

TabWhat it shows
OverviewTwo sections: 01 Text models (the LLM bias heatmap) and 02 Classic models (headline counts for your ML models). See Overview tab.
ML modelsEvery classic model on one list, with its source, task, latest version and drift, performance and fairness status. See ML models tab.
LLM runsThe runs behind the LLM bias heatmap.
BaselinesWhat each LLM series and each ML model version is compared against. See Baselines.
AlertsAlert rules for each tested LLM and each ML model. See Alerts.

What it checks on a classic ML model​

AreaWhat you get
DriftPSI on every feature and on the prediction, plus a KS test for numeric features and Jensen-Shannon distance for categorical ones. See Drift.
PerformanceAccuracy, precision, recall, F1 and ROC AUC for classifiers; MAE, RMSE and R² for regressions, each against the reference. See Performance and fairness.
FairnessFor each protected attribute, group-by-group rates checked against the four-fifths rule and an equalized odds limit (mean prediction and error gaps for regressions).
ExplanationsGlobal feature importance and how each feature's rank moved since the last version, plus the attributions behind one scored row. See Explanations and training.
LineageModel version, the training run that produced it and the dataset versions it learned from.

Your model stays with you​

QuilrAI never holds the model itself. It learns about a model from one of two sources:

  • Pipeline push: your pipeline registers the model and sends scored rows and metadata. See Connect from your pipeline.
  • MLflow: QuilrAI reads registered models, versions, runs, dataset inputs and feature importance from an MLflow tracking server or Databricks workspace you connect. Model files and datasets are never downloaded. See Connect an MLflow registry.

Exports from Drift & Bias contain check results only. Pushed rows, predictions and labels are never included.

What's new​

Drift & Bias changes are listed in the Red Teaming changelog.

Next steps​