radwa-f/BART-TweetDetox

BART Tweet Detoxification

Fine-tuned BART model for converting toxic text into neutral/polite versions while preserving semantic meaning.

Model Description

This model performs text-to-text transformation to detoxify toxic language. It takes potentially offensive or toxic text as input and generates a neutral, polite version while maintaining the original semantic content.

Key Features:

• Converts toxic tweets to neutral language
• Preserves semantic meaning
• Maintains fluency and naturalness
• Trained with prompt engineering techniques

Intended Uses

Primary Use Cases:

• Content moderation assistance
• Educational tools for communication
• Text sanitization for public platforms
• Research in NLP and text transformation

Out of Scope:

• Should not be used as sole content moderation tool
• Not intended for censorship purposes
• Requires human review for production use

How to Use

from transformers import pipeline

detoxifier = pipeline(
    "text-generation",
    model="radwa-f/bart-tweet-detoxification"
)

toxic_text = "You're such an idiot for thinking that!"
result = detoxifier(toxic_text, max_length=128)

print(result[0]['generated_text'])
# Output: "I respectfully disagree with your perspective on that."

Alternative usage with AutoModel:

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("radwa-f/bart-tweet-detoxification")
model = AutoModelForSeq2SeqLM.from_pretrained("radwa-f/bart-tweet-detoxification")

toxic_text = "You're such an idiot for thinking that!"
inputs = tokenizer(toxic_text, return_tensors="pt", max_length=128, truncation=True)
outputs = model.generate(**inputs, max_length=128, num_beams=5, early_stopping=True)

detoxified = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(detoxified)

Training Details

Training Data

• Dataset: ParaDetox
• Description: Parallel corpus of toxic and neutral text pairs
• Preprocessing: Text cleaning and normalization
• Language: English

Training Procedure

• Base Model: facebook/bart-large (or facebook/bart-base - confirm which you used)
• Task: Sequence-to-sequence generation for text style transfer
• Prompt Engineering: Applied controlled generation techniques
• Framework: PyTorch + Hugging Face Transformers

Hardware:

• GPU: Google Colab A100 GPU
• Training Time: 30 min

Hyperparameters

• Optimizer: AdamW
• Learning Rate: 2e-5
• Batch Size: 16
• Epochs: 3
• Max Length: 128 tokens

Evaluation

The model is evaluated on:

• Semantic Similarity: How well meaning is preserved
• Fluency: Naturalness of generated text
• Detoxification Rate: Reduction in toxic language
• BLEU Score: Translation quality metric
• ROUGE Score: Text overlap metric

Limitations

• May occasionally alter semantic meaning significantly
• Performance varies with highly context-dependent toxicity
• Training data biases may affect outputs
• Not 100% accurate - requires human oversight
• Works best with English text similar to training distribution

Ethical Considerations

This model is a research tool and should be used responsibly:

• Not a replacement for human moderation
• May reflect biases from training data
• Should be tested thoroughly before production use
• Users should be aware of limitations
• Consider privacy and consent when processing user-generated content

Model Creators

Radwa Fattouhi

• Final-year Software Engineering Student
• École Nationale des Sciences Appliquées (ENSA), El Jadida, Morocco
• Email: [email protected]
• LinkedIn: radwa-fattouhi
• GitHub: radwa-f

Amine Boktaya

• Final-year Software Engineering Student
• École Nationale des Sciences Appliquées (ENSA), El Jadida, Morocco
• Email: [email protected]
• LinkedIn: amine-boktaya
• GitHub: BoktayaAmine

Citation

If you use this model, please cite:

@misc{fattouhi2025detox,
  author = {Fattouhi, Radwa and Boktaya, Amine},
  title = {BART Tweet Detoxification},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/radwa-f/BART-TweetDetox}}
}

Model Card Authors

Radwa Fattouhi

• Final-year Software Engineering Student
• ENSA El Jadida, Morocco
• Contact: [email protected]

Related Work

Published research: AgriAlertX - SoftwareX (Elsevier)

---

Model card last updated: December 2025