Checklist ONIA 2026 County Stage, Grades 9–10 · Task 2
Fire Distinguisher
Profile vehicle sensor hazards, compute Mahalanobis anomaly scores and risk levels, and classify each reading as NORMAL, GAS_LEAK, OVERHEAT or FIRE.
The task
A car-safety company records one-second sensor readings from experimental vehicles (raw and averaged gas concentration, temperature, flame sensor, rates of change, sensor noise, battery level, speed, weather, sensor drift and ambient temperature) and needs a system that detects dangerous anomalies.
Subtask 1 asks for 12 statistics on train.csv (class percentages, mean temperature and mean gas concentration per class). Subtask 2 asks for a Mahalanobis anomaly score of each test reading relative to the NORMAL class, using the six numeric features with the strongest Spearman correlation to the encoded label, median imputation and standardisation fitted on train. Subtask 3 converts the scores into LOW/MEDIUM/HIGH using the 33rd and 66th percentiles of the training scores. Subtask 4 asks for the hazard class of each test reading.
Abridged and translated by SOTA from the official Romanian materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
train.csv(3,232 rows, 17 columns including label) andtest.csv(800 rows, without label). Archive of 370,388 bytes.- You submit
submission.csvwith subtaskID, datapointID and answer: 12 rows for subtask 1 and one row per test id for subtasks 2–4.- Scoring
- Subtask 1 (20): all values correct. Subtask 2 (15): proportional to the share of scores within ±0.05. Subtask 3 (15): proportional to the share of correct risk levels. Subtask 4 (50): weighted multi-class F1; 50 points at F1 ≥ 0.90, 0 below 0.60, linear in between.
- Rules
- On-site at county competition centres, in a prepared Docker/Jupyter environment (VS Code also available)
- Problem statements unlocked with a password handed out in the room
- Only the listed documentation may be consulted (NumPy, scikit-learn, OpenCV, Python, pandas, Matplotlib, seaborn, PyTorch, SciPy, CatBoost, Joblib)
- Format
- ONIA 2026 county stage, grades 9–10, 14 March 2026.