pjarbas312/OfensaBR-10k

OfensaBR Dataset

Dataset Description

Dataset Summary

OfensaBR is a Brazilian Portuguese dataset for offensive language detection, collected from social media Twitter/X. The balanced dataset contains 11k real-world comments and replies automatically classified using the Detoxify multilingual model.

This dataset was created to support:

• Content moderation systems for Brazilian Portuguese
• Offensive language detection models
• Toxicity classification research
• Test guardrails systems
• Social media analysis for PT-BR content

Supported Tasks and Leaderboards

• Text Classification: Binary classification of offensive vs. non-offensive content
• Toxicity Detection: Identification of toxic, harmful, or offensive language
• Content Moderation: Automated filtering of inappropriate content

Languages

• Portuguese (Brazil) - pt-BR

Ethical Considerations

Content Warning

⚠️ This dataset contains offensive language, including but not limited to:

• Profanity and vulgar language
• Insults and personal attacks
• Hate speech and discriminatory content
• Toxic and harmful expressions

Use responsibly and with appropriate safeguards.

Intended Use

✅ Appropriate Uses:

• Training content moderation systems
• Research on offensive language detection
• Developing safety tools for online platforms
• Academic study of online discourse
• Benchmarking NLP models

❌ Inappropriate Uses:

• Training systems to generate offensive content
• Targeting or harassing individuals
• Discrimination or profiling
• Surveillance without consent

Dataset Structure

Data Instances

Each instance contains:

{
  "id": 0,
  "text": "Sample text from social media",
  "category": "OFFENSIVE",
  "score": 0.87
}

Data Fields

• id (int): Unique identifier for the text
• text (string): The social media comment or reply text
• category (string): Classification label - either OFFENSIVE or NON_OFFENSIVE
• score (float): Confidence score from the classifier (0.0 to 1.0)

Dataset Creation

Curation Rationale

This dataset was created to address the lack of large-scale offensive language detection datasets for Brazilian Portuguese. Social media platforms contain real-world examples of offensive and non-offensive content, making them ideal sources for training content moderation systems.

Source Data

Initial Data Collection

Data was collected from public social media platforms with the following quality controls:

Quality Filters:

• Minimum text length: 10 characters
• Language detection: Portuguese only (using langdetect)
• Deduplication: Removes exact duplicates
• Text cleaning: URLs, mentions, and special characters normalized

Privacy Protection:

• ❌ No usernames collected
• ❌ No profile information stored
• ❌ No timestamps or metadata
• ✅ Only public comment text collected

Annotations

Annotation process

Automatic classification using Detoxify multilingual model:

• Model: unitary/multilingual-toxic-xlm-roberta
• Threshold: 0.5 (texts with score > 0.5 classified as OFFENSIVE)
• Categories:
  • OFFENSIVE: Toxic, harmful, threatening, insulting, or offensive content
  • NON_OFFENSIVE: Safe, neutral, or positive content

Who are the annotators?

Automated classification by Detoxify multilingual model. No human annotation was performed.

Personal and Sensitive Information

• ✅ All personal identifiers removed (usernames, handles, mentions)
• ✅ Only publicly available text collected
• ✅ No private or protected content included
• ✅ Complies with social media platforms' public data policies

Considerations for Using the Data

Social Impact of Dataset

Positive Impacts:

• Enables development of content moderation tools for Portuguese
• Supports research on offensive language detection
• Helps create safer online communities
• Provides benchmarks for PT-BR NLP models

Potential Risks:

• Dataset contains offensive language (by design)
• May reflect biases present in social media discourse
• Should not be used to target or discriminate against individuals
• Requires careful handling to avoid misuse

Discussion of Biases

Language Variation:

• Primarily Brazilian Portuguese (PT-BR)
• May include slang, informal language, and internet dialects
• Regional variations may be present

Classification Bias:

• Automated classification may have errors
• Model trained on specific toxicity definitions
• Cultural context of offensiveness may vary

Other Known Limitations

• Automatic Classification: No human verification of labels
• Threshold Sensitivity: Binary classification at 0.5 threshold may not capture nuance
• Context Loss: Text isolated from conversation thread
• Temporal Scope: Reflects language use at collection time
• Platform Changes: Social media dynamics evolve over time

Licensing Information

This dataset is licensed under the MIT License.

Citation Information

If you use this dataset in your research, please cite:

@dataset{ofensabr2025,
  title={OfensaBR: A Brazilian Portuguese Dataset for Offensive Language Detection},
  author={OfensaBR Contributors},
  year={2025},
  publisher={HuggingFace},
  howpublished={\url{https://github.com/PJarbas/OfensaBR-10k}}
}

Quality Metrics

Metric	Description
Language Purity	100% Portuguese (filtered with langdetect)
Deduplication	Exact duplicates removed
Min Length	All texts ≥ 10 characters
Privacy	No personal identifiers included

Contact

For questions, issues, or contributions:

• GitHub: https://github.com/PJarbas/OfensaBR-10k
• Issues: https://github.com/PJarbas/OfensaBR-10k/issues