---
license: cc-by-4.0
pretty_name: "IOAI 2026 – Double Agent Dilemma"
tags:
- ioai
- olympiad
- image-classification
viewer: false
---

# IOAI 2026 — Double Agent Dilemma

Official contest data for **Double Agent Dilemma**, task 4 (Day 2) of the [IOAI 2026](https://ioai-official.org) Individual Contest, held in Astana, Kazakhstan.

Two pretrained image classifiers — a ResNet18 (CNN) and a ViT-Tiny (Transformer) — both reach 100% accuracy on the provided images. The task exploits where the two architectures disagree.

The task statement, translations, baseline and grader live in the [IOAI-2026 GitHub repository](https://github.com/IOAI-official/IOAI-2026/tree/main/Individual-Contest/4_Double_Agent_Dilemma). This repository holds **data only**.

## Subsets

| Subset | Path | Contestants saw it | Role |
|---|---|---|---|
| Public – train | `public/train/` | yes | labelled images |
| Public – test | `public/test_public/` | yes | public evaluation split |
| Public – dry-run | `public/pretrain/`, `public/pretest/` | yes | small slices for pipeline validation |
| Private – Leaderboard A | `private/test_leaderboard_a/` | no | **live leaderboard** during the contest |
| Private – Leaderboard B | `private/test_leaderboard_b/` | no | **final ranking** |

Each split holds `images/` plus a `labels.json` mapping image index to class id.

## Credits

All of this data was extracted and filtered from ImageNet.

Proposed by **Tao Dajiang**; developed by **Tao Dajiang**, **Zhuldyz-Zhan Sagimbayev** and **Kamalkhan Artykbayev**. Published by the IOAI 2026 Scientific Committee under CC BY 4.0.