Facebook's Hateful Memes Challenge dataset (Kiela et al., 2020) contains 10,000 PNG meme images. The dataset is structured into training, development, and test splits, totaling 8,500, 500, 540, 1,000, and 2,000 entries respectively. This mirror was created by cs5242-hateful-memes for reproducibility of a CS5242 (NUS) submission.
Use Cases
- Train multimodal hate speech classifiers based on meme images and associated text.
- Benchmark model performance on seen and unseen splits for generalization testing.
- Study the intersection of visual and textual hateful content in social media memes.
- Develop preprocessing pipelines for image-text data pairs mentioned in the description.
Strengths
- Contains 10,000 PNG images, providing a substantial corpus for training.
- Includes structured splits (train, dev_seen, dev_unseen, test_seen, test_unseen) for robust evaluation.
- Based on a well-known benchmark dataset from Meta (Facebook), offering a recognized standard.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
- Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
- Source
- Mirror of the Facebook Hateful Memes Challenge dataset (Kiela et al., 2020).
- Collection Method
- Merges two existing mirrors of the original Meta release.
- Freshness
- Last updated 2026-04-24 12:28:47; freshness should be verified.