Principle:Allenai Open instruct Human Evaluation
| Knowledge Sources | |
|---|---|
| Domains | Evaluation, Annotation |
| Last Updated | 2026-02-07 02:00 GMT |
Overview
Principle of collecting and analyzing human preference judgments to evaluate and compare language model outputs through structured annotation protocols.
Description
Human evaluation provides ground-truth quality signals that automated metrics cannot fully capture. This principle involves presenting evaluators with model outputs in a controlled, bias-minimized setting (randomized presentation order), collecting structured judgments (acceptability ratings and pairwise preferences), and computing aggregate metrics with reliability measures. Key considerations include inter-annotator agreement (both strict exact-match and relaxed scoring where ties count as partial agreement), position bias mitigation through randomized presentation, and deduplication of annotations to handle re-evaluations.
Usage
Apply this principle when automated evaluation metrics are insufficient to distinguish model quality, when validating that automated metrics correlate with human preferences, or when producing human preference data for RLHF/DPO training.
Theoretical Basis
Human evaluation metrics follow established annotation science:
Acceptance Rate:
Win Rate (with clear/slight distinction):
Inter-Annotator Agreement (relaxed):
Pseudo-code Logic:
# Abstract human evaluation pipeline
for instance in evaluation_set:
shuffled = randomize_presentation_order(instance.completions)
judgments = collect_from_multiple_annotators(shuffled)
store(instance, judgments)
metrics = compute_acceptance_rates(judgments)
metrics += compute_win_rates(judgments)
metrics += compute_agreement(judgments)