Evaluation: Measuring AI Performance
“I poked it and it looked okay” is no longer a viable strategy for building AI applications. As models move into production, rigorous evaluation (Eval) becomes the most critical part of the development lifecycle.
Why Eval is Hard
Unlike traditional software where outputs are predictable, LLM outputs are:
- Non-deterministic: The same input can produce different outputs.
- Unstructured: Free-form text is harder to validate than a JSON response.
- Subjective: “Quality” or “Tone” are difficult to measure mathematically.
Types of AI Evaluation
- Deterministic Tests: Checking for specific keywords, valid JSON format, or exact matches in code.
- Model-Based Evals (LLM-as-a-Judge): Using a highly capable model (like GPT-4o) to grade the outputs of a smaller model based on a rubric.
- Benchmarks: Standardized tests like MMLU (knowledge), GSM8K (math), or HumanEval (coding).
- Human Evaluation: The gold standard, but the slowest and most expensive.
The Golden Dataset
The key to good evaluation is building a “Golden Dataset”—a curated list of inputs and their “correct” or “ideal” outputs. This allows developers to measure “Regression” (knowing if a change to a prompt actually made the model worse in some areas).
Metrics to Watch
- Faithfulness: Does the answer actually come from the provided context? (Crucial for RAG).
- Relevance: Does the answer actually address the user’s question?
- Perplexity: A mathematical measure of how “confused” the model is by the text.