CoreCognition: A Core Knowledge Benchmark for Multi-modal Large Language Models

Dataset Description

CoreCognition is a large-scale benchmark encompassing 12 core knowledge grounded in developmental cognitive science, designed to evaluate the fundamental core abilities of Multi-modal Large Language Models (MLLMs).

While MLLMs demonstrate impressive abilities over high-level perception and reasoning, their robustness in the wild remains limited, often falling short on tasks that are intuitive and effortless for humans. We examine the hypothesis that these deficiencies stem from the absence of core knowledge—rudimentary core abilities innate to humans.

This dataset contains 1,423 multimodal CoreCognition samples and 80 Concept Hacking questions with images/videos and questions, covering fundamental concepts like object permanence, spatial reasoning, counting, and other core abilities that emerge in human development.

🔗 Website: https://williamium3000.github.io/core-knowledge/
🔗 Paper: https://arxiv.org/abs/2410.10855
🔗 Github: https://github.com/williamium3000/core-knowledge

Formats

HuggingFace Preview - For browsing and exploration (visible in HuggingFace viewer, contains embedded 448*448-pixel image preview but no videos)

⚠️ Warning: this format is primarily for HuggingFace viewer; it DOES NOT contain full data.

Complete Dataset ZIP (Recommended) - Full data with all images and videos before resizing, 6.41GB

CoreCognition_20250622.zip
├── CoreCognition.csv          # Complete metadata CSV
└── media/                     # All images and videos
    ├── imagename1.png
    ├── imagename2.png
    ├── videoname1.mp4
    └── ...

Quick Start

Browse metadata and image preview in this huggingface repo
Download the complete dataset (6.41GB) by

from datasets import load_dataset

# this will downlo

27 个文件

浏览文件

数据集版权信息

本数据集的许可证为 Apache License 2.0。如有违反相关条款，请联系 WEHUB，我们将及时处理。查看许可证