Course · 5 chapters
Évaluation des LLM
Build runnable LLM evals you can trust: golden datasets, deterministic scorers, calibrated LLM judges, Inspect AI suites, and CI gates. 5 chapters, advanced, for engineers.
What you'll be able to do
- Build a runnable LLM eval from scratch
- Design judge rubrics that resist bias
- Calibrate LLM judges against humans
- Run eval suites with Inspect AI
- Gate bad merges with CI evals
- Set thresholds that survive flaky judges
What's inside
- 1Évaluation LLM : point de départ
Une orientation de 12 minutes sur le skill path LLM Evaluation — le chapitre passerelle, puis les trois couches (juges, suites, gates) qui transforment l'eval-par-intuition en discipline qui livre.
- 2Fondamentaux des evals : ta première eval LLM en 30 minutes
Arrête de vérifier les sorties au feeling. Construis une eval exécutable — dataset doré, scorer déterministe, juge LLM — et lis le résultat comme un ingénieur.
- 3LLM-as-Judge : rubriques, biais et fiabilité
Concevoir des juges qui résistent aux biais CALM, calibrer contre des humains et mériter une place dans ton gate CI.
- 4Inspect AI : suites d'évaluation en production à grande échelle
Crée, exécute et visualise des suites d'évaluation de niveau frontier avec le framework open-source de UK AISI.
- 5Gating par Evals en CI : Bloquer les Mauvais Merges
Intègre des evals par PR dans GitHub Actions, choisis des seuils qui résistent à la flakiness, et décide quand une gate appartient à main.
Frequently asked questions
- What will I learn in this LLM evaluation course?
- You build evaluation across three layers: a first runnable eval with a golden dataset and scorer, reliable LLM-as-judge rubrics calibrated against human ratings, and eval suites wired into CI as a merge gate. The path uses UK AISI's open-source Inspect AI framework and GitHub Actions.
- Who is this course for?
- It is for engineers building production AI features who need to test LLM outputs rigorously instead of checking them by vibes. The level is advanced, with a focus on software engineering and AI reliability.
- Do I need to code to take this course?
- Yes. This is a hands-on engineering path that involves writing eval scripts, configuring the Inspect AI framework, and setting up GitHub Actions workflows, so comfort with code and CI is expected.
- How long is the course and is there a certificate?
- The path has 5 chapters totaling about 100 minutes, starting with a 12-minute orientation. On completion you earn an AI Academy by Anthropos certificate.
- Is this course free?
- No, this is a paid skill path included with an AI Academy by Anthropos subscription.
Earn a certificate
Complete all chapters to receive your certificate of completion.