HateSpeechMix

HateSpeechMix is a research-oriented normalization of four multilingual hate/offensive-speech resources. It contains 187,315 rows across Amharic, Chinese, English, Hindi, and Turkish with deterministic train and test splits.

All published rows contain text. English and Turkish use the authors' official 2024 full-text v2 release rather than the older ID-only export.

Schema

Field	Meaning
text	Source text
label	Three-way ClassLabel: 0 = neutral, 1 = offensive, 2 = hate
label_text	Human-readable form of label
fine_label	Best available normalized subtype; preserves Hindi profane separately
language	ISO 639-1 language code
source	Stable source-corpus identifier
source_id	Identifier within the source corpus
source_label	Original source label(s), retained for auditability
original_split	Original split, or unsplit where the build created one
topic	Source topic when available
target	Source target type when available
text_available	Whether text contains usable text

Label normalization

Source	Unified mapping
English/Turkish	Normal → neutral; Offensive → offensive; Hate → hate
Amharic	Normal → neutral; Offensive → offensive; Hate → hate
Hindi	NONE → neutral; OFFN or PRFN → offensive; HATE → hate
Chinese	Safe → neutral; every offensive/attack category → offensive

Chinese has no unified hate examples. Its test-only individual/group distinction remains in source_label and target; mapping group attacks to hate would add an interpretation absent from the source's main binary annotation. Hindi PRFN remains visible as fine_label = "profane" and in source_label even though its unified label is offensive.

Split policy

• HASOC Hindi: preserve official train and test splits.
• COLD Chinese: preserve test; merge upstream dev into train.
• Amharic RANLP 2023: preserve test; merge upstream dev into train.
• English/Turkish v2: create a deterministic 90/10 split, stratified by language, topic, and original three-way label. The source used 10-fold cross-validation and does not publish a fixed split.

After assigning splits, the build removes training rows whose exact (source, text) pair also occurs in test. Test rows are retained. This removes 56 leaked training rows (54 Chinese, one Hindi, and one Amharic), including nine whose unified label conflicts with the test copy. Within-split duplicates are retained and reported rather than silently resolved because some have conflicting annotations.

Rows are kept in source order and are not globally shuffled. Shuffle explicitly during training.

Composition

Language	Published rows	Upstream rows	Text available
Amharic (am)	15,000	15,001	15,000
Chinese (zh)	37,426	37,480	37,426
English (en)	68,597	68,597	68,597
Hindi (hi)	5,982	5,983	5,982
Turkish (tr)	60,310	60,310	60,310
Total	187,315	187,371	187,315

Exact split and label counts are recorded in metadata/build_summary.json. Source revisions and checksums are recorded in metadata/source_manifest.json.

The official English/Turkish TSV contains bare carriage returns inside some tweet texts. The builder normalizes these formatting characters before parsing, validates required fields and categorical values, and reports any dropped records under source_quality in the build summary. The pinned release retains all 128,907 English/Turkish rows.

Licensing and redistribution

This compilation does not have a single permissive license:

• English/Turkish: CC BY-NC-SA 4.0.
• Chinese COLDataset: Apache-2.0.
• HASOC 2019 Hindi: the supplied archive states no license; the official distribution is password protected and requests registration.
• Amharic RANLP 2023: neither the Hugging Face card nor its linked source repository states a license.

license: other is intentional. Users must comply with each source's terms. In particular, the absence of a stated license must not be interpreted as permission for unrestricted redistribution or commercial use.

Reproducibility

The builder pins every public source revision and verifies SHA-256 checksums. It requires the locally supplied hindi_dataset.zip because HASOC uses registered access.

python -m venv .venv
.venv/bin/python -m pip install -r hatespeechmix/requirements.txt
.venv/bin/python hatespeechmix/scripts/build_dataset.py
.venv/bin/python -m pytest hatespeechmix/tests

The build writes Hub-ready Parquet files to data/ and a local Arrow-backed DatasetDict to local_dataset/ (ignored by Git).

Sources and citations

Please cite the original resources rather than this mechanical compilation.

HASOC 2019 Hindi

@inproceedings{mandl2019hasoc,
  title={Overview of the HASOC track at FIRE 2019: Hate Speech and Offensive Content Identification in Indo-European Languages},
  author={Mandl, Thomas and Modha, Sandip and Majumder, Prasenjit and others},
  booktitle={Working Notes of FIRE 2019},
  year={2019}
}

English/Turkish

@inproceedings{toraman2022large,
  title={Large-Scale Hate Speech Detection with Cross-Domain Transfer},
  author={Toraman, Cagri and Sahinuc, Furkan and Yilmaz, Eyup Halit},
  booktitle={Proceedings of LREC 2022},
  pages={2215--2225},
  year={2022}
}

Chinese COLDataset

@inproceedings{deng2022cold,
  title={COLD: A Benchmark for Chinese Offensive Language Detection},
  author={Deng, Jiawen and Zhou, Jingyan and Sun, Hao and Mi, Fei and Huang, Minlie},
  booktitle={Proceedings of EMNLP 2022},
  pages={11580--11599},
  year={2022}
}

Amharic RANLP 2023

@inproceedings{ayele2023exploring,
  title={Exploring Amharic Hate Speech Data Collection and Classification Approaches},
  author={Ayele, Abinew Ali and Yimam, Seid Muhie and Belay, Tadesse Destaw and Asfaw, Tesfa and Biemann, Chris},
  booktitle={Proceedings of RANLP 2023},
  pages={49--59},
  year={2023}
}

Ethical considerations

The text contains slurs, harassment, and identity-directed abuse. Labels reflect different annotation guidelines, collection periods, sampling strategies, and cultural contexts; they should not be treated as interchangeable ground truth. Models trained on the mixture may reproduce annotation bias and may perform unevenly across languages and target groups.