Checklist Moldova AI Olympiad 2026 Simulation #1 (selection round) · Task 1
The Whispering Codru
Multi-label toxicity classification of online comments into six categories, preceded by two label-counting subtasks.
The task
In a Moldovan folk-tale setting, harmful messages appear on a village board and an AI guardian must recognise harmful speech. The data are online comments, some manually annotated with six binary labels: toxic, severe_toxic, obscene, threat, insult and identity_hate; a comment may have several labels or none.
Subtask 1 (10 pts) counts the training comments with each label equal to 1. Subtask 2 (20 pts) counts training comments with exactly 0, 1, …, 6 active labels. Subtask 3 (70 pts) trains a model that outputs six probabilities for every test comment.
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
train.csv(comments with numeric IDs starting from 1 and the six labels);test.csv(comments without labels, IDs continuing after the training IDs).- You submit
- One CSV with subtaskID, datapointID, answer: category name and count (subtask 1);
exactly_k_labelsand count (subtask 2); test comment id and six space-separated probabilities in the order toxicsevere_toxicobscene threat insultidentity_hate(subtask 3). - Scoring
- Subtasks 1 and 2: exact counts, each correct datapoint earns an equal share. Subtask 3: mean column-wise ROC AUC; 0 pts below 0.88, linear between 10 and 69 pts for 0.88 ≤ AUC < 0.96, 70 pts at ≥ 0.96.
- Rules
- Python, Google Colab, Kaggle Notebook and standard data-science/ML libraries are allowed.
- No collaboration and no help from other people or AI assistants; internet only for documentation.
- Zoom session with screen sharing, camera, microphone and phone camera, plus OBS screen recording, for the whole round.
- Submission limit 50 (platform settings).
- Format
- Simulation #1, a preparatory round of the National AI Olympiad, 16 May 2026, online on Nitro AI Judge (3 hours from 10:00 local time; platform window 07:00–10:25 UTC), proctored via Zoom and OBS; its results count 30% towards selecting the 12 finalist teams.