DigniFy Multimodal Multilingual Hate Speech Datasets

DigniFy Multimodal Multilingual Hate Speech Datasets

Data Contents and Provenance

Text

english_preprocessed_data.csv: Filtered subset of the “Measuring Hate Speech” dataset from Hugging Face (ucberkeley-dlab/measuring-hate-speech). test_hinglish.parquet and train_hinglish.parquet: Train/validation/test splits of the IPD-Text-Hinglish dataset from Hugging Face (dj-dawgs-ipd/IPD-Text-Hinglish). 

Images

We curated and scraped subsets of images depicting various hateful or offensive gestures and symbols. 

image_symbols.zip:

  - swastika/  - finger_gun_to_the_head/  - middle_finger/  - cut_throat_gesture/  - slanted_eyes/

Third‑party datasets linked but not hosted:

- Roboflow “Nazi Swastika vs. Hindu Swastika” (rm-yoq1a/swast)

- Roboflow “Chinese Eyes” Object Detection Dataset by DaeHo (daeho/5-szwgp) 

- Roboflow "Middle Finger" Dataset by El Hareketi Deneme: (deneme-yz/el-hareketi-deneme) 

Audio

All raw video files were obtained from the HateMM repository. We provide here only the preprocessing files for the spectrograms and features used in the audio model. 

License

This collection is licensed under CC-BY 4.0 for all derivatives. Third‑party assets remain under their original licenses (RF/RM for Getty, CC-BY for Hugging Face, etc.). 

Citation

If you use this dataset, please cite:

bibtex@dataset{bhathawala2025dignify,  author       = {Tirath Bhathawala},  title        = {DigniFy Multimodal Multilingual Hate Speech Datasets},  publisher    = {Zenodo},  year         = {2025},  doi          = {10.5281/zenodo.15637274}}

For questions or access to raw data not included here, contact the corresponding author.