Checklist ONIA 2026 National Stage · Task 1
Gender biases shall not pass: Test-time debiasing of pretrained embeddings
Steer CLIP image embeddings so that they no longer encode gender while remaining useful for unknown downstream tasks.
The task
Because retraining large pretrained models is prohibitively expensive, changes must be made without retraining, for instance by model steering. The contestant must design a method that removes gender bias from the image-encoder activations of OpenAI's pretrained CLIP model.
Given 15,000 training embeddings with binary gender labels and 15,000 unlabelled test embeddings, the contestant outputs debiased versions of both matrices. The steered embeddings are evaluated on four undisclosed tasks for task performance and for gender bias of the predictions.
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
main.pyprovidingtrain_image_embeddings(15000 × 512),gender_labels(15000),test_image_embeddings(15000 × 512) andgenerate_submission(). Archive of 57,201,116 bytes.- You submit
submission.pklproduced bygenerate_submissionfrom the two debiased (15000 × 512) matrices, uploaded inside a .zip archive.- Scoring
- Two scores over four hidden tasks, (i) task performance (higher is better) and (ii) bias (closer to 0 is better); final ranking by their harmonic mean, ties broken by the debiasing score. The platform lists subtasks 'Păstrare performanță' (accuracy) and 'Debiasare' (KL divergence).
- Rules
- On-site at the National University of Science and Technology POLITEHNICA Bucharest
- No internet access; offline resources only (NLTK, spaCy
en_core_web_sm, scikit-learn, Gensim glove-wiki-gigaword-50, fastText without pretrained models, distilbert-base-uncased) - Downloading models or data is forbidden; allowed documentation listed in the round description
- Problem statements unlocked with a password handed out in the room
- Format
- ONIA 2026 national stage, single contest day, 18 April 2026 (06:00–10:20 UTC on the platform); same problems for grades 9–10 and 11–12.