ALLMAcademy

D4

Four Eval Types You Layer, Not Pick

12 min
  1. 01Distinguish golden, regression, adversarial, LLM-judge, and human evals by the question each answers
  2. 02Design regression evals that survive nondeterminism using tolerant scorers and aggregate gates
  3. 03Name three LLM-judge biases and the mitigations that bring it closer to human labels
  4. 04Explain the method triangle: where to spend cheap automated, judge, and scarce human labels
Narration — Four Eval Types You Layer, Not Pick
0:00 / 0:00

Five eval types, five questions

topic 17

You don’t choose one eval type — you layer them, because each answers a different question:

  • Golden set — curated, version-controlled inputs + reference outputs. Is this representative behavior correct?
  • Regression — re-run the golden set on every prompt/model/retrieval change as a CI gate. Did this change make things worse?
  • Adversarial / red-team — injection, jailbreaks, OOD, edge cases. Is it safe under attack?
  • LLM-as-judge — pointwise or pairwise scoring at scale. How good, cheaply, at volume?
  • Human — slow, expensive ground truth for subjective/high-stakes calls.
Where each eval type plugs into the loopInteractive · 2D

Production traces feed sampled failures into the golden set; the golden set powers the regression gate that guards the next release. Adversarial, judge, and human evals layer on top.

Loading diagram…
Text description

Production traces feed sampled failures into the golden set; the golden set powers the regression gate that guards the next release. Adversarial, judge, and human evals layer on top.

Regression under nondeterminism

Never assert exact string equality. Even at temperature 0, output varies via batching, hardware, MoE routing, and silent provider updates. Use tolerant scorers — schema/regex checks, semantic similarity, or an LLM judge — and gate on aggregate score deltas, not single examples.

The method triangle

Spend scarce human labels to calibrate cheaper graders

A pyramid: cheap automated checks on every run, LLM-judge on samples, human review on a small high-value or judge-disagreement subset. You layer methods so scarce human labels calibrate the cheaper ones (track inter-annotator agreement). Well-built judges reach roughly 80% agreement with humans on some tasks (Zheng et al., 2023) — task-dependent, so validate before trusting.

Common MisconceptionA high golden-set score means the system is safe.

CorrectionTypical-case and adversarial evals answer different questions — a high golden score says nothing about injection resistance.

Common MisconceptionAn LLM judge gives objective ground truth.

CorrectionAn unvalidated judge is just a second untested model — calibrate it against human labels and average mirrored positions first.

Production failure modes

  • Golden-set leakage/overfitting: examples seen in training inflate scores.
  • Stale golden set: green evals while real users complain.
  • Judge position/verbosity bias letting a worse-but-longer answer win.
  • Self-preference inflation: a judge favoring outputs from its own model family.
  • Tiny non-representative set making the CI gate noisy and untrustworthy.

Retrieval Practice

Check one idea at a time

Question 1 of 2

An exact-match grader marks a correct reply wrong because it said "Sure!" where the reference said "Of course!". Best fix?