LLM Evaluation

概要

LLM Evaluation provides a robust framework for systematically measuring the performance of language model applications through automated metrics, human-in-the-loop workflows, and advanced LLM-as-judge patterns. It enables developers to move beyond subjective testing by providing implementation guidance for classic NLP metrics like BLEU and ROUGE alongside modern embedding-based scores and RAG evaluations. With built-in patterns for A/B testing, regression detection, and benchmarking, this skill ensures that model updates and prompt changes result in measurable improvements rather than unexpected regressions.

主な機能

• Automated metric implementation for text generation, classification, and RAG retrieval.
• Statistical A/B testing to measure performance gains and significance.
• Human evaluation frameworks with annotation guidelines and agreement calculations.
• LLM-as-judge patterns for pointwise scoring and pairwise model comparisons.
• 23,139 GitHub stars
• Automated regression detection to prevent performance drops during deployment.

ユースケース

• Detecting and preventing performance regressions in CI/CD pipelines.
• Validating the groundedness and factual accuracy of RAG systems.
• Comparing model performance and prompt iterations with quantitative data.