Saved locally. It will merge when you sign in.
| Task type | Two 90-minute rounds on the same day: an AI fundamentals paper, then an applied problem in Google Colab. |
|---|---|
| Dataset | Round 1 is self-contained; Round 2 supplies a dataset with the notebook. |
| I/O format | Round 1: 25 multiple-choice and 10 fill-in-the-blank answers. Round 2: a trained model plus the code that produced it. |
| Metric | Round 1 is 100 marks. Round 2 is judged on model performance metrics and code quality. |
| Limits | 90 minutes per round. Python throughout. Round 2 runs in Google Colab. |
The 2025 National Qualifier is the most useful past paper available for CAIO, and the organisers publish it: the Round 1 problem set as a PDF, the Round 2 Colab notebook with its test dataset, and two practice tests for each round.
Ninety minutes, 100 marks:
The material spans mathematical logic, basic coding, algorithmic thinking and AI foundations. Questions may involve equations, hand reasoning about algorithms, and intuitive understanding of machine learning principles. Everything is in Python where code appears.
The fill-in-the-blank half carries half the marks across ten questions, so each one is worth two and a half multiple-choice questions. They reward exact answers rather than recognition, which changes how you should allocate time within the round.
Ninety minutes, one practical problem in Google Colab. You are given a dataset and expected to preprocess it, apply a suitable machine learning algorithm, and submit both the trained model and your code.
Marks come from two places: model performance metrics, and code quality. That second component is easy to overlook and it means a fast, unreadable notebook does not score full marks even if its accuracy is good.
The organisers note that scoring around 35% on either practice test puts you in a competitive position. This paper is not designed to be finished comfortably, and treating a mid-30s score as a failure will mislead you about your standing.
Sit both rounds under real conditions - 90 minutes each, no interruptions, no searching - before reading any solutions. The published practice tests exist so you can do this twice. Test 2 is noted as matching the actual difficulty more closely than Test 1.
Two problems in the shape of Round 1, to try before opening the paper.
A recommender system has to suggest a film to someone who has just created an account. Which of these best describes the cold-start problem?
Stuck? There are 2 hints for this problem.
Hints cost no marks, but the real paper has none. Try first.
Collaborative filtering scores a user by similarity to other users' histories. A new account has no history, so there is nothing to compare, which is the cold-start problem. Content-based features or a popularity prior are the usual first answers.
Select an answer
Two users rated the same three films. User A gave and user B gave . What is the cosine similarity between them, to one decimal place?
To one decimal place, . Round 1's fill-in-the-blank questions say how many decimal places they want; match it exactly, because the marking is by exact answer.