LLMs, GenAI, RAG, agents, evaluation · insight-1585
Debias a judge-scored pass rate with a small human-labelled sample instead of reporting it raw
- Evidence type
- documented fact
- Difficulty
- hard
- Roles
- Data scientist, Applied AI
- Industries
- cross-industry
Summary: situation, action, reason, boundary
Situation
A judge scores thousands of production conversations cheaply, but you also hold human labels for a small subset.
Action
On the labelled subset estimate the average human-minus-judge difference, add it to the judge's mean on the large set, and build the interval from both samples. Prediction-powered inference gives a principled version; applying it to LLM-judge scores is this course's adaptation, not a claim of the original paper.
Reason
The small gold sample anchors the level and the large judged sample reduces variance, so the combined interval stays valid without assuming the judge is accurate.
Boundary
Estimating a rate or mean over large unlabelled traffic with a random labelled subset.
Symptom or risk
The judge's pass rate is a biased estimate of the human-defined rate, and a bare interval on it ignores that bias.
Exceptions
If the labelled subset is not a random sample of the same traffic, the correction is itself biased.
Trade-offs
A better judge gives a narrower interval; a poor judge gives a wide one, which is honest.
Pitfalls
Labelling only the cases the judge found hard; sample at random.
Example
Hypothetical: judge passes 82% of 2,000 items; on 150 labelled items humans pass 72% versus the judge's 80%. Corrected estimate 74%; if about 19% of the labelled items disagree, the 95% interval is roughly plus or minus 7 points.
Lessons and concepts
Claims and their sources
Prediction-powered inference provides valid confidence intervals for quantities such as means from a small gold-labelled set plus a large set of model predictions, without assumptions about the model. Using it with LLM-judge scores is this course's adaptation.