IOAI 2025 · Individual Contest (set)
The 2nd International Olympiad in AI (Beijing, August 2–9, 2025) ran a two-day Individual Contest with 6 tasks spanning ML, NLP, CV, and concept reasoning. This set links the three most representative walkthroughs — including the existing chicken-counting CV problem.
Official source.
ioai-official.org/2025-tasks ·
IOAI-official/IOAI-2025 (GitHub) ·
Hugging Face dataset.
Six tasks: Radar, Chicken Counting, Concepts, Restroom Icon Matching, Antique Painting Authentication,
Pixel Efficiency.
Round metadata
| Year | 2025 (2nd edition) |
|---|---|
| Host | Beijing, China · 2–9 August 2025 |
| Round | Individual Contest (two days, three tasks per day) |
| Tasks (6) | Radar (ML), Chicken Counting (CV), Concepts (NLP), Restroom (CV), Antique (CV), Pixel (optimisation) |
| Duration | 2 days · ~6–8 contest hours per day [verify in official schedule] |
| Hardware | Per-contestant Linux box · Python 3.12.7 · pinned dependencies |
| Allowed | Any open-weights model; per-task notebook supplied with baseline cells |
| Scoring | Per-task metric scaled to 0–100, summed across the 6 tasks [illustrative weighting] |
Featured tasks
CV
Chicken counting
Crowd-style density-map regression on aerial chicken-farm images. (Existing walkthrough.)
Open walkthrough →
ML
Radar
Classify radar returns / detect objects from radar signal cubes. A signal-processing-flavoured ML task: features matter more than the model.
Open walkthrough →
NLP
Concepts
Reason about and classify abstract concepts from text. Tests whether you can prompt-engineer a small open-source LLM into reliable structured answers.
Open walkthrough →
How to use this set
- Pull the official notebooks from Individual-Contest/ and run them locally on the supplied datasets. The baseline cells are part of the problem — read them before writing anything.
- Sit two of the tasks back-to-back under a 6-hour clock to feel real Day-1 fatigue.
- For each task, write down your hypothesis ("this is a density-regression problem") before reading the official solution notebook. The diff between your hypothesis and the official one is where you learn.
- The three tasks not deeply walked through here (Restroom, Antique, Pixel) are all CV / optimisation flavoured — try them once you've finished the featured three; the toolchain transfers.