Evaluating LLM applications and agents

How to tell whether an application got better after a change: evaluation datasets, LLMs as judges and their biases, agent reliability, RAG evaluation, statistical significance and observability in production.

 

Loading…

01The whole topic

Take the whole "Evaluating LLM applications and agents" topic

Every question of the topic — one per fact, from easy to hard. An honest check of the whole topic rather than of a single section.

02

Topic sections

#QUIZQUESTIONSDIFFICULTYSTATUS
6.1Evals and datasetsWhat to build an evaluation set from, labeling, linking metrics to business goals
6.2LLM as a judgePosition and length biases, self-preference, rubrics
6.3Evaluating agentsPass@k and Pass^k, τ²-bench, verifiers, error attribution
6.4RAG evaluationContext recall, faithfulness, answer correctness
6.5Statistics and productionSignificance of differences, observability, evaluation on live trafficSections: Statistical significance · Observability and tracing
6.6Benchmarks and their limitsWhat public benchmarks are for, Ranking by pairwise comparisons, An unexpected win in Elo

Ranking needs a sign-in

A ranked attempt is available after signing in with Telegram: that way one person is one participant in the ranking.

How the quizzes and the fair ranking work

01Modes

Practice
No timer and no limit on the number of attempts. After every answer you see the correct option and an explanation. It does not affect the ranking. Available without signing in.
Ranked attempt
Time for each question is limited; correctness and explanations come only after the last question. The result goes into the ranking. Requires signing in with Telegram.
Diagnostic
32 questions, about 20 minutes: for each of the eight topics one easy, two medium and one hard question, new every time. It is a check for yourself: no sign-in and no ranking, and at the end — a knowledge map and a list of what is worth brushing up on. Topics and their sections count toward the ranking.

02Time and answers

  • The server keeps the time, not the browser: from 30 to 90 seconds per question depending on difficulty and format. Reloading the page does not reset the timer — the question comes back with the same time.
  • Questions come in different formats: pick one option, mark all correct ones, put steps in order, find the faulty line.
  • If time runs out, the question counts as skipped. A small allowance for network delay is built in.
  • You cannot go back to a previous question. Sending an answer again changes nothing — the first one counts.
  • Number keys select elements, Enter sends the answer. In practice you can open a hint.
  • The "Not sure about the answer" mark does not affect the points, but a correct answer with it goes into the review as a topic worth brushing up on.
  • You can change the site language at any time, even in the middle of an attempt: the question and the timer stay the same.

03Limits on ranked attempts

  • One ranked attempt per quiz per day. Days are counted in UTC — a new day starts at 00:00 UTC.
  • Switching to another tab or minimizing the window during a ranked question counts it as skipped.
  • Leaving the quiz early: the remaining questions count as skipped, and you can retry this quiz the next day.

04How points are counted

  • Every question has its own difficulty rating, refined by the answers of all participants. A correct answer to a hard question gives more points, a mistake on an easy one costs more.
  • A skip and "time is up" count as mistakes. The "not sure" mark does not affect the points.
  • Every participant has an overall rating and a separate rating for each topic. The starting rating is 1000.
  • The ranking is the same for the Russian and English versions of the site: the questions and their difficulty are shared.
  • The "Week" and "Month" tables show how many points a participant gained or lost over the calendar week or month. "All time" is the current rating.
  • Sometimes a control question is added to a ranked attempt. It does not count toward the score.

05A fair ranking

  • One Telegram account is one participant. When you sign in, we receive only your name and username.
  • An attempt clearly taken by a program rather than a person does not count toward the ranking.
  • Fast answers alone are not penalised: strong participants answer fast too.