ioai-2026-double-agent-dilemma
收藏资源简介:
IOAI 2026 – Double Agent Dilemma 是国际人工智能奥林匹克(IOAI)2026年个人竞赛第四天(任务4)的官方数据集,于哈萨克斯坦阿斯塔纳举办。该数据集包含两个预训练的图像分类器——ResNet18(CNN)和ViT-Tiny(Transformer),它们在提供的图像上均达到100%的准确率。任务旨在利用两种架构之间的分歧点。数据来源于ImageNet的提取和过滤。数据集分为多个子集:公开训练集(public/train/)、公开测试集(public/test_public/)、公开试运行集(public/pretrain/和public/pretest/)、私有排行榜A集(private/test_leaderboard_a/)和私有排行榜B集(private/test_leaderboard_b/)。每个子集包含图像文件夹(images/)和标签文件(labels.json),其中标签文件将图像索引映射到类别ID。该数据集适用于图像分类任务,特别关注不同神经网络架构的决策差异。数据集采用CC BY 4.0许可证。
IOAI 2026 – Double Agent Dilemma is the official dataset for Day 4 (Task 4) of the Individual Competition of the International Olympiad in Artificial Intelligence (IOAI) 2026, held in Astana, Kazakhstan. The dataset contains two pre-trained image classifiers—ResNet18 (CNN) and ViT-Tiny (Transformer)—both achieving 100% accuracy on the provided images. The task aims to exploit the divergence between the two architectures. The data is sourced from ImageNet extractions and filtering. The dataset is divided into multiple subsets: public training set (public/train/), public test set (public/test_public/), public holdout sets (public/pretrain/ and public/pretest/), private leaderboard set A (private/test_leaderboard_a/), and private leaderboard set B (private/test_leaderboard_b/). Each subset contains an images folder (images/) and a labels file (labels.json), where the labels file maps image indices to class IDs. This dataset is suitable for image classification tasks, with a special focus on decision differences across different neural network architectures. The dataset is licensed under CC BY 4.0.
IOAI 2026 – Double Agent Dilemma 数据集概述
基本信息
- 数据集名称:IOAI 2026 – Double Agent Dilemma
- 许可证:CC BY 4.0
- 数据类型:图像分类
- 标签:ioai、olympiad、image-classification
任务背景 该数据集是 IOAI 2026 个人赛第 2 天任务 4 的官方竞赛数据,比赛在哈萨克斯坦阿斯塔纳举行。任务围绕两个预训练图像分类器——ResNet18(CNN)与 ViT-Tiny(Transformer)——在提供的图像上均达到 100% 准确率,但两模型在预测不一致之处展开。
数据集结构 数据集仅包含数据,任务说明、翻译、基线和评分器均存放在 IOAI-2026 GitHub 仓库中。包含以下子集:
| 子集 | 路径 | 参赛者可见 | 用途 |
|---|---|---|---|
| Public – train | public/train/ | 是 | 带标签图像 |
| Public – test | public/test_public/ | 是 | 公开评估划分 |
| Public – dry-run | public/pretrain/, public/pretest/ | 是 | 用于流程验证的小样本 |
| Private – Leaderboard A | private/test_leaderboard_a/ | 否 | 比赛期间的实时排行榜 |
| Private – Leaderboard B | private/test_leaderboard_b/ | 否 | 最终排名 |
每个划分均包含 images/ 文件夹及 labels.json 文件,后者将图像索引映射到类别 ID。
数据来源与致谢 所有数据均从 ImageNet 中提取并经过筛选。该任务由 Tao Dajiang 提出,并由 Tao Dajiang、Zhuldyz-Zhan Sagimbayev 和 Kamalkhan Artykbayev 开发,由 IOAI 2026 科学委员会以 CC BY 4.0 许可发布。




