← All problems

🇧🇬 Bulgaria team selection 2025, IOAI Team Selection

CyrillicCross: Recovering Visual Features Across Languages

Learn a projector that maps corrupted (dropout plus circular shift) Bulgarian RoBERTa token embeddings onto the CLIP text embeddings of the English prompt.

To hand in
A trained model and its notebook
Hardware
Needs a GPU (free Colab works)
Data size
1.51 MB
Data type
📝 Text
Estimated difficulty
★★★☆☆
Solution
See a solution ↗ · best opened after you have tried

Get ready on the platform

The chapters of the learning path that prepare for this problem.

These chapters are picked from the solution: showing them may hint at the approach.

›How to tackle it
  1. Read the statement. What to hand in, the metric that scores you, the time and hardware limits.
  2. Explore the data. Size, format, missing values, how the labels are spread.
  3. Score yourself first. Keep part of the training data aside to measure your work before submitting.
  4. Start simple, then improve. A first solution that works end to end, then one change at a time.
  5. Compare with a solution. Once you have searched, open it and see what you would have done differently.

Unofficial copy: the organisers' original may differ.

The statement, data and solutions belong to their authors; we only link to them.

Are you the author or organiser? Ask for a correction or removal