AiActivity/ToxicDataset

---
license: mit
task_categories:
- text-classification
- text-generation
- other
language:
- en
tags:
- toxic-content
- hate-speech
- content-moderation
- abuse-detection
- nlp
- safety
- moderation
- offensive-language
pretty_name: Comprehensive Toxic Content Dataset
size_categories: 1M<n<10M
---

# Comprehensive Toxic Content Dataset

## Dataset Description

This dataset contains **1,000,000 synthetically generated records** of toxic, abusive, harmful, and offensive content designed for training content moderation systems and hate speech detection models.

### Dataset Summary

This comprehensive dataset includes multiple categories of toxic content:
- **Toxic** content (insults, derogatory terms)
- **Abusive** language patterns
- **Gender bias** statements
- **Dangerous/threatening** content
- **Harmful slang** and abbreviations
- **Racist** content patterns
- **Homophobic** content
- **Religious bias** statements
- **Disability bias** content
- **Mixed** category combinations

### Supported Tasks

- **Text Classification**: Multi-class classification of toxic content types
- **Severity Detection**: Classification of content severity (low, medium, high, extreme)
- **Content Moderation**: Training moderation filters and safety systems
- **Hate Speech Detection**: Identifying hate speech patterns
- **Abuse Detection**: Detecting abusive language online

### Languages

The dataset is primarily in **English (en)**, with patterns based on real-world English-language toxic content from social media platforms.

## Dataset Structure

### Data Fields

Each record contains the following fields:

- **id** (`int`): Unique identifier for the record
- **content** (`string`): The toxic content text
- **category** (`string`): Category of toxic content (toxic, abusive, gender_bias, dangerous, harmful_slang, racist, homophobic, religious_bias, disability_bias, mixed)
- **severity** (`string`): Severity level (low, medium, high, extreme)
- **timestamp** (`string`): ISO format timestamp
- **metadata** (`dict`): Additional metadata including:
- `language`: Language code (en)
- `type`: Content type (text)
- `source`: Source identifier (generated)
- `flagged`: Boolean flag indicating toxic content

### Data Splits

The dataset can be split into train/validation/test sets. Recommended splits:
- **Train**: 80% (800,000 records)
- **Validation**: 10% (100,000 records)
- **Test**: 10% (100,000 records)

## Dataset Statistics

### Category Distribution
- Toxic: ~10%
- Abusive: ~10%
- Gender Bias: ~10%
- Dangerous: ~10%
- Harmful Slang: ~10%
- Racist: ~10%
- Homophobic: ~10%
- Religious Bias: ~10%
- Disability Bias: ~10%
- Mixed: ~10%

### Severity Distribution
- Low: ~25%
- Medium: ~25%
- High: ~25%
- Extreme: ~25%

## Dataset Creation

### Source Data

This dataset is synthetically generated based on patterns and vocabulary from:

1. **Academic Research**:
- Davidson et al. (2017): Hate Speech Detection on Twitter
- Waseem & Hovy (2016): Twitter hate speech patterns
- Founta et al. (2018): Large-scale abusive behavior
- Zampieri et al. (2019): Offensive language identification

2. **Public Datasets**:
- Jigsaw Unintended Bias in Toxicity Classification (2M+ comments)
- Hate Speech and Offensive Language Dataset (25K tweets)
- Toxic Comment Classification Challenge (160K+ comments)
- HateXplain Dataset (20K+ posts)
- OLID Dataset (14K tweets)

3. **Real-World Sources**:
- Hatebase.org lexicon
- Jigsaw Perspective API patterns
- Documented patterns from social media platforms

### Annotation Process

- **Pattern-based generation**: Uses comprehensive word lists and pattern templates
- **Validation**: All records validated for required fields and content quality
- **Balanced distribution**: Ensures balanced representation across categories
- **Realistic variations**: Includes leetspeak, character repetition, punctuation variations

### Personal and Sensitive Information

This dataset contains **synthetic toxic content** generated for research purposes. It does not contain real personal information or actual harmful content from individuals. All content is algorithmically generated based on documented patterns.

## Considerations for Using the Data

### Ethical Considerations

⚠️ **WARNING**: This dataset contains toxic, abusive, harmful, and offensive content.

**Intended Use**:
- Training content moderation systems
- Building safety filters and detection models
- Academic research on online toxicity
- Developing hate speech detection algorithms
- Educational purposes for understanding toxic content patterns

**NOT Intended For**:
- Harassing individuals or groups
- Creating harmful content
- Targeting marginalized communities
- Any malicious purposes

### Limitations

1. **Synthetic Content**: All content is algorithmically generated, not real user-generated content
2. **English Only**: Primarily English language patterns
3. **Pattern-Based**: May not capture all nuances of real-world toxic content
4. **Bias**: Patterns based on documented research may reflect biases in source data

### Bias and Fairness

- The dataset is designed to be balanced across categories
- Patterns are based on documented research and public datasets
- Users should be aware of potential biases in source materials
- Regular evaluation and bias testing recommended for production models

## Citation

```bibtex
@dataset{toxic_content_dataset_2024,
title={Comprehensive Toxic Content Dataset for Moderation Training},
author={Dataset Generator},
year={2024},
url={https://huggingface.co/datasets/[USERNAME]/toxic-content-dataset},
note={Generated for research and content moderation purposes only}
}
```

### Source Citations

```bibtex
@inproceedings{davidson2017automated,
title={Automated Hate Speech Detection and the Problem of Offensive Language},
author={Davidson, Thomas and Warmsley, Dana and Macy, Michael and Weber, Ingmar},
booktitle={Proceedings of the 11th International AAAI Conference on Web and Social Media},
year={2017},
pages={512-515}
}

@inproceedings{waseem2016hateful,
title={Hateful Symbols or Hateful People? Predictive Features for Hate Speech Detection on Twitter},
author={Waseem, Zeerak and Hovy, Dirk},
booktitle={Proceedings of the NAACL Student Research Workshop},
year={2016},
pages={88--93}
}
```

## Dataset Card Contact

For questions or concerns about this dataset, please refer to the repository issues or contact the maintainers.

## License

This dataset is released under the **MIT License**. See LICENSE file for details.

## Acknowledgments

This dataset is based on patterns and vocabulary from:
- Academic research on hate speech and toxic content detection
- Public datasets from Jigsaw, Davidson et al., and other researchers
- Hatebase.org lexicon
- Documented patterns from social media platforms

We thank all researchers and organizations who have contributed to understanding and detecting toxic content online.