AI & LLM Evals / Scenario

Cleaning the Golden Set

The eval score is stable, and wrong. The golden dataset behind it is a mess. Audit it and decide what to fix before you trust another number.

Difficulty
Medium
Format
Scenario
Points
150
Estimate
14 min

// MISSION BRIEF

Your Mission

A support-bot eval reports a steady 0.82, but production complaints keep rising. The metric is not the problem, the golden dataset is. You have a sample of it and the maintenance history.

Audit the dataset hygiene and prescribe the fixes.

// FIRST CONTACT

Battle teaser

First artifact

Golden set sample (rows)

Which dataset-hygiene problems are present? (Select all that apply.)

  1. ANear-duplicate rows (1, 2, 3) over-weight one intent
  2. BSevere class imbalance (47/60 password-reset) skews the score
  3. CNo refusal / out-of-scope cases, so the eval never tests boundaries
  4. DA stale label (row 4 still says 14 days; policy is now 30)
Answers, scoring, hints, and the full battle stay sealed.

// SKILL TAGS

ai-evalsgolden-datasetsleakagedata-hygiene