AKCIT/ToxSyn-PT

Dataset Summary

ToxSyn-PT is a large-scale synthetic dataset designed for fine-grained hate speech detection in Brazilian Portuguese. It comprises 53,274 sentences equally balanced between toxic and non-toxic labels, covering nine legally protected minority groups (including Black, Women, LGBTQIA+, Native Brazilian, Muslim, Jewish, and Elderly).

Unlike most existing datasets that only capture hostile mentions, ToxSyn-PT systematically includes non-toxic counterexamples (benign text discussing minorities). This helps models learn to distinguish genuine hate speech from harmless discussions, reducing bias and preventing mutual generalization failures common in social-media domains.

Dataset Structure

Data Instances

A typical instance includes a synthetic sentence, its toxicity label, the specific minority group it mentions, and the rhetorical strategy used.

Example:

{
  "text": "A contribuição da comunidade muçulmana para a nossa cultura é inegável.",
  "toxicity": "Non-toxic",
  "target_group": "Muslim",
  "discourse_type": "Praise/Endorsement"
}

Data Fields

• —`text` (string): The generated sentence in Portuguese.
• —`toxicity` (string): Binary classification indicating whether the text is Toxic or Non-toxic.
• —`target_group` (string): The specific minority group discussed in the text.
• —`discourse_type` (string): The rhetorical strategy of the text (e.g., sarcasm, dehumanization, endorsement).

Dataset Creation

To ensure broad linguistic coverage and class balance, ToxSyn-PT was created using a controllable four-stage LLM pipeline:

1. 1.Compact Seed: Manual curation of high-quality initial examples.
2. 2.Few-Shot Expansion: Using instruction-tuned LLMs to generate new varied instances.
3. 3.Paraphrase Augmentation: Expanding the syntactic variety of the generated text.
4. 4.Enrichment: Adding neutral and supportive texts to curb model overfitting to group-specific keywords.

Ethical Considerations

ToxSyn-PT contains highly toxic and offensive language by design, as it is intended to train robust safety classifiers. However, all data is synthetically generated and does not correspond to real individuals. The dataset should be used strictly for research in harm mitigation, hate speech detection, and bias reduction in NLP systems.

Citation

If you use this dataset, please cite the following paper:

@inproceedings{brito-etal-2026-toxsyn,
  title = {ToxSyn-PT: A Synthetic Fine-Grained Dataset of Minority-Targeted Toxic Language in Portuguese},
  author = {Brito, Iago Alves and Dollis, Julia Soares and Farber, Fernanda Bufon and Fernandes, Diogo and Filho, Arlindo R. Galvão},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference (LREC 2026)},
  month = {May},
  year = {2026},
  pages = {3908--3920},
  address = {Palma, Mallorca, Spain},
  publisher = {European Language Resources Association (ELRA)},
  editor = {Piperidis, Stelios and Bel, Núria and van den Heuvel, Henk and Ide, Nancy and Krek, Simon and Toral, Antonio},
  doi = {10.63317/3ne367tx8hvj},
}