MemeBias-5: A Multiclass Dataset for Detecting Bias and Sentiment in Internet Memes

MemeBias-5 is a curated dataset of 2,243 internet memes categorized into five distinct classes — Hateful (361), Misogynistic (458), Racist (338), Normal (497), and Sarcastic (589). Each meme represents real-world social media content encompassing humor, bias, and sentiment cues expressed through text and visuals.

This dataset is designed for research in multimodal bias detection, sentiment classification, hate-speech analysis, and explainable AI (XAI). MemeBias-5 can be used to train and evaluate models that integrate visual and textual understanding, including CNNs, Vision Transformers, and Vision-Language Transformers (e.g., CLIP, ViLT, BLIP).

The dataset provides a benchmark for exploring fairness, interpretability, and robustness in social media AI systems, encouraging the development of responsible AI models that understand meme context beyond surface sentiment.

File structure:
MemeBias-5/
├── Hateful/ (361 images)
├── Misogynistic/ (458 images)
├── Racist/ (338 images)
├── Normal/ (497 images)
└── Sarcastic/ (589 images)
Data type: Image (.jpg, .png)
Total files: 2,243
Total size: (Add final compressed size, e.g., ~200 MB)

Intended Use:
-This dataset supports research in:
-Bias and hate-speech detection
-Meme sentiment and sarcasm classification
-Multimodal representation learning
-Explainable and ethical AI

Ethical Considerations:
All data were collected from publicly available sources for educational and research purposes. Offensive content is included solely for analysis of bias and harmful language. Users are advised to handle the dataset responsibly, respecting ethical guidelines and local regulations.