EngineersofAI LogoCoursesBlogResearch LabAI LettersThe LabCode BankInteractive 3DKodrEarnestJobs
Skip to main content
Interactive 3D/Human Evaluation - Annotation and Agreement
Model Output
Reinforcement learning from human feedback (RLHF) fine-tunes a language model by training a reward model on human preference judgments, then using PPO to optimize the language model to maximize that reward.
Rater Annotations (3 raters, 1-5 scale)
Raterhelp…fact…safe…clar…
Rater 14554
Rater 24553
Rater 33454
Avg3.74.75.03.7
Inter-Rater Agreement (Fleiss Kappa)
helpfulnessκ=0.94 Strong
factualityκ=0.94 Strong
safetyκ=1.00 Strong
clarityκ=0.94 Strong
Controls
Number of raters3
2345
Criteria
Human evaluation scores each output on multiple criteria. Fleiss Kappa above 0.8 = strong agreement; below 0.6 = raters need calibration. Hover a rater row to highlight their scores. Compare outputs A/B/C to see which model is most factual.

Human Evaluation - Annotation and Agreement - Interactive Visualization

Human evaluation of LLM outputs involves multiple raters scoring each response on criteria like helpfulness, factuality, safety, and clarity on a 1-5 scale. Fleiss kappa measures how consistently raters agree - above 0.8 is strong, below 0.6 means raters need calibration. Disagreements are most common on subjective criteria like helpfulness, and reveal where annotation guidelines need tightening.

  • Fleiss kappa above 0.8 indicates strong inter-rater agreement; below 0.6 requires guideline revision
  • Safety scores show the highest inter-rater agreement; helpfulness the lowest due to subjectivity
  • Rater calibration sessions resolve systematic disagreements before large-scale annotation begins
  • Aggregated scores across 3-5 raters are statistically more reliable than any single rater judgment

Part of the EngineersOfAI Interactive 3D - free interactive visualizations covering every major concept in machine learning and AI engineering. Hover any element for a plain-English explanation. No code required.