Discord

Checklist OFIA 2026 Stage 1 (qualification round) · Task 1

Where Is Spy-C?

French title: Où est Spy-C ?

Predict which of five animal agents a mission belongs to from tabular data, then locate that animal in the corresponding test image.

  • Tabular
  • Vision
  • Multi-class tabular classification combined with object localisation in images
  • French original

The task

The first qualifying round of OFIA 2026 is titled 'Où est Spy-C ?' ('Where is Spy-C?'). According to the dataset description, the data are synthetic and were designed for OFIA 2026; they combine tabular descriptive features with visual elements such as images of animals. Each row of the tabular data describes a mission (meeting-point coordinates, radio coordinates, mission type, agent status, speciality, language level, contact frequency, mission duration, operations budget and a discretion score), and the training labels give the identity of the agent (spy_identity).

The identified agents of the 'Spies Group' are five animals — classes 0 to 4: beaver (castor), fox (renard), pigeon, rabbit (lapin) and raccoon (raton laveur). Five sprite images per animal are supplied as clues to each agent's appearance. The 1,000 test images are hosted on Hugging Face; according to the starter notebook they measure 1024×1024, 2048×2048 or 4096×4096 pixels.

According to the starter notebook published with the dataset, the identity predicted for row N of X_test determines which animal must be found in test image N (for example, a prediction of 1 on the first row means that the fox must be located in test_0000.png); the submission gives, for each test row, the predicted identity and the (x, y) pixel position of the animal in the image.

Abridged and translated by SOTA from the official French materials. The official statement has the exact rules, and it wins wherever this summary differs.

At a glance

You get
X_train.csv (2000 rows, 12 columns: coord_rencontre_x, coord_rencontre_y, coord_radio_x, coord_radio_y, type_mission, statut_agent, specialite, niveau_langue, frequence_contact, duree_mission, budget_ops, score_discretion); y_train.csv (target spy_identity); X_test.csv (1000 rows); espions/{castor,renard,pigeon,lapin,raton-laveur}/sprite_5.pngsprite_9.png; 1000 test images test_0000.pngtest_0999.png on Hugging Face (OlympiadesIA/OFIA2026); exemples/exemple_chargement_images.py, exemples/generer_exemple_soumission.py and exemples/exemple_soumission.csv.
You submit
CSV with columns id, spy_identity, spy_position_x, spy_position_y — one row per test sample (format taken from the starter notebook).
Rules
  • Individual participation only; one Kaggle account per person.
  • Open to French nationals or pupils studying in France who are under 20 on 2 August 2026 and have not started higher education; parental consent required under Kaggle rules.
  • Up to 100 submissions per day; up to 2 submissions selected for final evaluation.
  • AI assistants (ChatGPT, Mistral, Gemini, Claude, …) may be used to think about or write code, but candidates must be able to explain their ideas and code; third-party services called from the code (e.g. the OpenAI API) are not allowed.
  • The code that produced the final submissions must be uploaded to the organisers; solutions are checked by code review.
  • Discussing solutions with anyone, and sharing the statement or data outside Kaggle, are prohibited during the competition.
Format
OFIA 2026 Stage 1 (qualification), online on Kaggle, 1–28 Feb 2026 (deadline 28 Feb 2026 at 23:59, Paris time according to the starter notebook). Up to 100 best-scoring solutions on the private leaderboard qualify for Stage 2, a 3-hour online round on 21 Mar 2026.

Details

Year
2026, Online (Kaggle)
Round
Stage 1 (qualification round) · Task 1
Language
French
License
CC BY-NC-ND 4.0 (Attribution-NonCommercial-NoDerivatives 4.0 International), licence of the Kaggle dataset timmor/ofia2026; the competition rules restrict use of the statement and data to the competition, as stated by the source