Model Evaluation (Evals)
What is it?
Model evaluation — 'evals' — is how you measure whether an AI model or feature actually does its job well, using repeatable tests instead of gut feel.
Explain like I'm 5
Why was it created?
AI output is open-ended and easy to judge by vibes. Evals were adopted to catch regressions and compare models objectively as prompts, models, and data change.
Where is it used?
- Comparing models or prompts
- Regression testing before deploy
- Measuring accuracy, safety, cost
- Monitoring quality in production
Why should developers care?
Without evals you can't tell if a prompt tweak or model swap helped or hurt, so they're essential for shipping and improving AI features safely.
How does it work?
You assemble a dataset of inputs with expected outcomes or scoring criteria, run the model, and score results — via exact match, rules, human raters, or an 'LLM-as-judge' that grades responses. Aggregated scores reveal whether a change is an improvement.
Real-world example
Before switching models, you run both against 200 saved support questions and compare accuracy, refusals, and cost to decide objectively.
Common use cases
- A/B testing prompts and models
- Pre-deploy regression checks
- Safety and bias testing
- Tracking quality over time
Advantages
- Objective, repeatable comparisons
- Catches regressions early
- Guides prompt/model choices
- Builds confidence to ship
Disadvantages
- Good test sets take effort to build
- Open-ended tasks are hard to score
- LLM-as-judge has its own biases
- Benchmarks can be gamed or leak into training
When should you use it?
Any time you change a prompt, model, or pipeline and need to know if quality improved — and continuously in production.
When should you avoid it?
Never truly avoid it; for a throwaway prototype a quick manual check may suffice temporarily.
Alternatives
Related terms
Interview questions
Beginner
- Why evaluate an AI model instead of judging by feel?
- What is a simple way to score model outputs?
Intermediate
- What is 'LLM-as-judge' and its risks?
- Why build a task-specific eval set instead of using public benchmarks?
Senior
- How do you guard against benchmark contamination?
- How would you design evals for an open-ended generation task?
Common misconceptions
- "Public benchmark scores prove it works for me" — your task and data may differ, and benchmarks can leak into training.
- "Evals are only for training" — they matter just as much for prompts and production monitoring.
Fun facts
- 'LLM-as-judge' uses one model to grade another's answers at scale.
- A small, honest, task-specific eval set often beats a big generic benchmark.
Timeline
- 2020s — Task-specific evals become core to LLM development
- 2023 — LLM-as-judge popularized for scalable grading
Learning resources
Quick summary
Model evaluation measures whether an AI model or feature works using repeatable tests, so you can compare options and catch regressions objectively.
Cheat sheet
- Repeatable tests, not vibes
- Exact-match, rules, human, or LLM-judge
- Catches regressions on changes
- Task-specific sets beat generic benchmarks