Checklist ROAI 2026 County Stage, Grades 9–10 · Task 2
The Orient-Inator vs. Perry the Platypus
Using only a frozen three-class orientation classifier, recover 4-way and 8-way image rotations and detect the classifier's own mistakes.
The task
A cheap pretrained multinomial logistic-regression 'oracle' for MNIST-like images only outputs probabilities for Upright (0°), Upside-Down (180°) and Sideways (90° or 270°). It may not be retrained, and no new image classifier may be fitted; contestants may only transform the inputs (skimage.transform.rotate) and query predict_proba.
Subtask 1 asks for the exact rotation among 0°, 90°, 180° and 270°. Subtask 2 asks for the exact rotation among eight angles in 45° steps, including diagonals the oracle has never seen. Subtask 3 asks, without ground truth, for a binary flag indicating whether the oracle's original prediction on an image is wrong.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
- Test archive with the pretrained model (
doof_orackle.pkl) and a separate test subset for each subtask;starter_kit.pywith loading and processing code. - You submit
- CSV with subtaskID (1–3), datapointID and answer (angle class 0–3, angle class 0–7, or mistake flag 0/1).
- Scoring
- Subtask 1 (30 points): accuracy. Subtask 2 (40 points): accuracy. Subtask 3 (30 points): Matthews correlation coefficient.
- Rules
- No retraining and no new image classifiers; only rotations and the oracle's
predict_probamay be used - On-site at partner county centres, in the prepared local Jupyter environment
- Only the packages listed in the round rules may be used
- Mobile phones and other electronic or information sources are forbidden in the room
- No retraining and no new image classifiers; only rotations and the oracle's
- Format
- ROAI 2026 county stage, grades 9–10, 8 March 2026.