Skip to main content

Regression Metrics

Regression Metrics use an LLM-as-Judge to compare two responses directionally — a new response containing your latest changes (the candidate) against a known-good past response (the baseline), resolved from a tagged run.

Instead of an absolute score, they look for Shifts in Behavior, outputting a -1, 0, or 1:

Output ScoreMeaning
-1The Baseline (old version) is better or exhibits more of the trait.
0Tie. The old and new versions are effectively the same regarding this trait.
1The Candidate (new version) is better or exhibits more of the trait.

Aggregated across the dataset, these become a win / loss / tie rate that you gate with max_loss_rate or min_win_rate in a target's regression block.

Regression vs Pairwise

Regression is directional — it asks "did we get better or worse than the baseline?". Its sibling, Pairwise (Fairness) Metrics, is symmetric — it compares two peer responses for divergence with no notion of a baseline. They are separate axes with separate metric libraries; a regression metric cannot be used in a pairwise block, and vice versa.

Available Regression Metrics

All regression metrics share the same variables: {{input}}, {{output_baseline}}, and {{output_candidate}}.

general-quality

A holistic assessment. Which response is better overall based on accuracy, helpfulness, and clarity?

decisiveness

Which response is more decisive and less cautious (less hedging, fewer "maybes")?

refusal

Identifies if one model refused the user's request (e.g., "I cannot answer that") while the other answered it.

verbosity

Which response is longer or more verbose?

instruction-following

Which response followed the system instructions and constraints more rigorously?

formality

Which response is more formal, academic, or professional in tone?

ambiguity-handling

When faced with a vague input, which response correctly asked for clarification instead of wrongly guessing the intent?

apologetic

Identifies which response is more apologetic or subservient.

brand-voice

Which response sounds more aligned with a professional, innovative, and customer-centric brand voice?

empathy

Which response demonstrates superior emotional intelligence and empathy towards the user's situation?

simplicity

Which response explains complex concepts more simply, avoiding unnecessary technical jargon?

actionability

Which response provides clearer next steps or calls to action to the user?