102 models · $0.00 inference · NVIDIA Build API
Learn to evaluate
AI agents.
An eleven-lesson course on building evaluations, from a first dataset to a release gate. Every lesson ends in a lab you run against real models — no install, no notebook, and no bill.
Start lesson 0111 lessons · about 2.1 hours
Ticket triage · accuracy by category
example from lesson 08
| shipping | billing | returns | account | escalate | |
|---|---|---|---|---|---|
| llama-3.3-70b | 0.96 | 0.94 | 0.70 | 0.92 | 0.90 |
| nemotron-super-49b | 0.94 | 0.92 | 0.68 | 0.90 | 0.88 |
| qwen2.5-7b | 0.90 | 0.86 | 0.62 | 0.84 | 0.80 |
| mistral-small-24b | 0.88 | 0.84 | 0.58 | 0.82 | 0.78 |
Every model scores lowest on returns. When models from four different labs share a weakness, the cause is usually the labels rather than the models — lesson 04 covers how to tell.
What you build
One worked example throughout: a support-triage agent, 50 labelled tickets, 13 of them deliberately debatable.