README.md

metadata

license: cc-by-4.0
language:
  - en

MoralCLIP Dataset

Dataset for training and evaluating multimodal models on moral content understanding based on Moral Foundations Theory (MFT).

This dataset contains 15,000 image-text pairs annotated with moral labels across 5 moral foundations:

• Care / Harm
• Fairness / Cheating
• Loyalty / Betrayal
• Respect / Subversion
• Sanctity / Degradation

Dataset Structure

Each entry contains:

• id: Unique identifier
• source_dataset: Source of the image (imagenet, laion, or smid)
• caption: Primary caption describing the image
• captions: List of 10 alternative generated captions (5 for LAION examples)
• moral_labels: List of applicable moral foundations

Splits:

• Train: 13,500 samples
• Validation: 750 samples
• Test: 750 samples

Downloading Images

The dataset references images from three sources. Follow the instructions below to obtain the images:

1. SMID (Social-Moral Image Database)

Download from: https://osf.io/2rqad/

2. LAION-400M

Available here: https://laion.ai/blog/laion-400-open-dataset/

3.ImageNet

Requires registration

1. Register at: https://image-net.org/download.php
2. Download ILSVRC2012 validation set
3. Images are named like: ILSVRC2012_val_00012789.JPEG Note: ImageNet requires agreeing to their terms of use for research purposes.

License

This dataset follows the licenses of its source datasets:

SMID: Check OSF page ImageNet: Research use only, requires registration LAION: Creative Commons license