Hugging Face's logo

Datasets:
Shuwan
/
cadet-datasets

Tasks:
Text Classification
Modalities:
Tabular
Text
Formats:
json
Languages:
English
Size:
100K - 1M
ArXiv:
Tags:
hate-speech
toxicity
safety
DOI:
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
Community
1

Request access to IE-HSC

The datasets used in this research contain sensitive information and are made available for academic, non-commercial research purposes only.

Thank you for your interest in our paper, "Causality Guided Representation Learning for Cross-Style Hate Speech Detection," TheWebConf 2026
The datasets used in this research contain sensitive information and are made available for academic, non-commercial research purposes only. To request access, please complete this form.
We are committed to upholding the highest ethical standards for data privacy and responsible research. All requests will be reviewed by our team, and we will contact you via the email provided upon approval.

Log in or Sign Up to review the conditions and access this dataset content.

Implicit-Explicit Hate Speech Corpus (IE-HSC)

GitHub Paper Conference

This dataset accompanies the paper "Causality Guided Representation Learning for Cross-Style Hate Speech Detection" (TheWebConf/WWW 2026).

IE-HSC contains hateful, harassing, and otherwise offensive language, including identity-based attacks and slurs. This content may be distressing and can cause harm if misused.

Use this dataset responsibly. Follow your organization's ethics and safety policies, minimize exposure to raw text, and avoid redistributing offensive examples unless necessary for research.

Dataset Summary

Implicit-Explicit Hate Speech Corpus (IE-HSC) is a dataset of implicit/explicit style hate speech samples from multiple sources. It is designed for research on hate speech detection, especially cross-style generalization between implicit and explicit hate speech.

IE-HSC is based on four open-sourced datasets:

Dataset Access

IE-HSC is intended for academic, non-commercial research use only. This dataset is provided as a gated dataset on the Hugging Face Hub; request access on the dataset page.

Dataset Overview

Dataset Split Total Count Hate Ratio Style Ratio
AbuseEval explicit_test 589 98.1% 17.8%
AbuseEval explicit_train 2,353 98.8% 17.4%
AbuseEval implicit_test 2,060 14.0% 0.5%
AbuseEval implicit_train 8,238 14.7% 0.6%
DynaHate explicit_test 2,206 65.6% 100.0%
DynaHate explicit_train 8,823 64.1% 100.0%
DynaHate implicit_test 6,021 49.6% 0.0%
DynaHate implicit_train 24,084 50.2% 0.0%
Implicit-Hate-Corpus explicit_test 301 70.8% 64.5%
Implicit-Hate-Corpus explicit_train 1,201 72.8% 64.9%
Implicit-Hate-Corpus implicit_test 4,028 35.5% 3.5%
Implicit-Hate-Corpus implicit_train 16,110 36.2% 4.0%
IsHate explicit_test 2,732 86.3% 51.5%
IsHate explicit_train 10,928 86.0% 52.8%
IsHate implicit_test 9,673 67.4% 0.6%
IsHate implicit_train 38,689 67.2% 0.6%
  • Hate ratio: fraction of examples with hate_label = 1.
  • Style ratio: fraction of examples with style = 1 (derived from avg, see below).

Data Structure

Each example includes:

  • text_id: ID for each text sample (string)
  • text: Input text content (string)
  • hate_label: Binary hate label (0=non-hate, 1=hate)
  • avg: Average Perspective API toxicity score (float, 0.0-1.0)
  • style: Binary toxicity derived from avg (0=non-toxic, 1=toxic)
  • true_style: Style from file naming (0=implicit, 1=explicit)
  • target: Target demographic group (string)
  • target_conf: Confidence score for target annotation (float, 0.0-1.0)

Provenance and Processing

  1. Label standardization

    • Columns label, hateful_layer, or hate_label are normalized to hate_label.

  2. How avg is computed

    • Perspective API scores for TOXICITY, SEVERE_TOXICITY, IDENTITY_ATTACK, INSULT, PROFANITY, and THREAT are averaged per example and stored as avg.

  3. Toxicity and style

    • If avg is missing, default to 0.0.
    • avg >= 0.4 ⇒ style = 1 (toxic/explicit); avg < 0.4 ⇒ style = 0 (non-toxic/implicit).
    • true_style is taken strictly from the dataset file name (explicit_* ⇒ 1, implicit_* ⇒ 0).
    • For DynaHate, style equals true_style.

Intended Use

IE-HSC is intended for research and educational use only (non-commercial), such as:

  • Studying implicit vs. explicit hate speech
  • Training and evaluating hate/toxicity classifiers
  • Benchmarking robustness, bias mitigation, and safety interventions

Out-of-Scope / Misuse

  • Using the data to generate, amplify, or target hateful content
  • Deploying models trained on IE-HSC in user-facing or high-stakes settings without additional validation, safeguards, and human oversight
  • Treating target labels or toxicity scores as ground truth about individuals or communities

Ethical Considerations

  • Labels and target annotations may be noisy and reflect biases in source datasets and automated scoring tools.
  • If you share examples (papers, demos, model cards), include content warnings and avoid quoting slurs or targeting language unless strictly necessary.

License

IE-HSC is licensed under the CC BY-NC-SA 4.0 license. Because IE-HSC is derived from multiple sources, please also review and comply with the licenses and terms of the underlying datasets.

Citation

If you use IE-HSC in your research, please cite:

@article{zhao2025causality,
  title={Causality Guided Representation Learning for Cross-Style Hate Speech Detection},
  author={Zhao, Chengshuai and Wan, Shu and Sheth, Paras and Patwa, Karan and Candan, K Sel{\c{c}}uk and Liu, Huan},
  journal={arXiv preprint arXiv:2510.07707},
  year={2025}
}

@dataset{cadet_dataset_2026,
  title={cadet-datasets},
  url={https://huggingface.co/datasets/Shuwan/cadet-datasets},
  DOI={10.57967/HF/7615},
  publisher={Hugging Face},
  author={Chengshuai Zhao, Shu Wan, Paras Sheth, Karan Patwa, K. Selçuk Candan, Huan Liu}, year={2026}
}

@software{shu_2026_18331976,
  author       = {Shu},
  title        = {Shu-Wan/cadet: Initial release (v1.0.0)},
  month        = jan,
  year         = 2026,
  publisher    = {Zenodo},
  version      = {v1.0.0},
  doi          = {10.5281/zenodo.18331976},
  url          = {https://doi.org/10.5281/zenodo.18331976},
  swhid        = {swh:1:dir:ceb3fc11806a24207cb0e43fd92b9b6f106f4cc2
                   ;origin=https://doi.org/10.5281/zenodo.18331975;vi
                   sit=swh:1:snp:402265cabd7f46dfbeeadf2a29ed9ed9a173
                   48a8;anchor=swh:1:rel:014aced8c618a22ad8ffd659e14f
                   8c8e29e11440;path=Shu-Wan-cadet-678b13c
                  },
}
Downloads last month
45

Collection including Shuwan/cadet-datasets

Paper for Shuwan/cadet-datasets