liveBook · Manning

Canary set

As explained in AI Model Evaluation

A canary set is a fixed evaluation dataset used to detect changes in evaluator behavior or other evaluation-pipeline components.

A canary set is a fixed evaluation dataset used to detect changes in evaluator behavior or other evaluation-pipeline components. Re-running it against a stored baseline helps distinguish a judge change from a genuine regression in the system being evaluated. section 9.8 section 10.6.2

Use a canary set as part of a repeatable validation and triage loop:

  • Select representative cases that reflect the evaluation task, including easy, ambiguous, erroneous, and difficult examples. section 9.8
  • Collect human ratings or gold labels, then compare judge outputs with those references before scaling evaluation. section 9.8 section 9.11
  • Lock the prompt, rubric, judge model, and baseline results; rerun the fixed set whenever a pipeline component changes. section 9.8 section 10.6.2
  • When scores move, check dataset changes, judge versions, prompts, rubrics, and stability before declaring a model regression. section 10.6.2

Recommendation-explanation canary set

Include explanations that are accurate, hallucinated, irrelevant, or persuasive but unsupported by the user’s history. Compare the judge’s results with trusted labels, inspect disagreements, and rerun the same cases after changing the prompt or judge model. section 9.8

A canary set is not a substitute for validation, stability testing, or human judgment. It is a recurring regression signal that makes evaluator drift visible and helps target further investigation. section 10.7 section 9.11

How AI Model Evaluation covers it

How deep it goes
Defines selection, labeling, baseline locking, reruns, and disagreement checks, but defers detailed stability-test procedures.
Examples
Uses recommendation explanations spanning accurate, hallucinated, irrelevant, and ungrounded persuasive outputs.
What it assumes
Assumes a defined evaluation task, rubric, judge, and trusted human ratings or gold labels for comparison.
What you can do after
Diagnose evaluator drift and separate judge changes from genuine model regressions before spending engineering effort on the wrong cause.
How it connects
Connects validation, gold labels, human ratings, regression testing, prompt versioning, and stability testing.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Model Evaluation. The section links point to the parts of the book it draws on.