bytop/propaganda-detection-historical-context

Propaganda Detection with Historical Context

This model is a fine-tuned version of IDA-SERICS/PropagandaDetection specifically trained to detect historical context manipulation and geopolitical propaganda.

Model Description

• Base Model: IDA-SERICS/PropagandaDetection (DistilBERT-based)
• Fine-tuned on: 168 examples of historical context propaganda
• Training Accuracy: 100% validation accuracy
• Performance Improvement: +66.7 percentage points on geopolitical propaganda detection

Key Capabilities

The model excels at detecting:

• Geopolitical Framing: Biased presentation of conflicts ("Israel's war", "special military operation")
• Genocide Denial: Language that minimizes or denies documented genocides
• War Crimes Euphemisms: Sanitized language for documented violations ("collateral damage", "surgical strikes")
• False Equivalence: Creating false moral equivalence between different actions
• Victim Blaming: Language that blames victims of historical atrocities
• Historical Revisionism: Attempts to rewrite established historical facts

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("bytop-ai/propaganda-detection-historical-context")
model = AutoModelForSequenceClassification.from_pretrained("bytop-ai/propaganda-detection-historical-context")

# Create pipeline
classifier = pipeline("text-classification", model=model, tokenizer=tokenizer)

# Analyze text
result = classifier("Israel's war against Hamas continues")
print(f"Propaganda: {result[0]['label']} ({result[0]['score']:.3f})")

Performance

Model Type	Historical Context Detection	Overall Accuracy
Base Model	16.7%	54.5%
This Model	83.3%	90.9%
Improvement	+66.7 points	+36.4 points

Training Data

The model was fine-tuned on a carefully curated dataset of 168 examples including:

• 100 historical context propaganda examples
• 68 traditional propaganda examples
• Balanced representation of different propaganda techniques
• Examples covering Israel-Palestine, Ukraine-Russia, and other geopolitical contexts

Ethical Considerations

This model is designed for educational purposes to help users:

• Recognize propaganda techniques in media
• Understand how language can obscure accountability
• Develop critical thinking about historical framing
• Identify bias in geopolitical reporting

The model's classifications are based on documented facts and international law, not political positions.

Limitations

• Trained primarily on English text
• Focused on contemporary geopolitical contexts
• May not generalize to all historical periods
• Requires careful interpretation in sensitive contexts

Citation

@misc{bytop-propaganda-historical-context-2024,
  title={Fine-tuned Propaganda Detection with Historical Context},
  author={BytoP.ai},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/bytop-ai/propaganda-detection-historical-context}
}

Model Architecture

• Architecture: DistilBERT for Sequence Classification
• Parameters: ~67M (same as base model)
• Max Sequence Length: 512 tokens
• Labels: 2 (NO_PROPAGANDA: 0, PROPAGANDA: 1)

Training Details

• Training Epochs: 5
• Learning Rate: 2e-5
• Batch Size: 4
• Validation Strategy: 20% held-out
• Early Stopping: 3 patience
• Final Validation Accuracy: 100%

Example Detections

Propaganda (Detected)

• "Israel's war against Hamas continues as they defend themselves" → PROPAGANDA (99.7%)
• "Russia's special military operation to denazify Ukraine" → PROPAGANDA (95.9%)
• "The alleged genocide is just Hamas propaganda" → PROPAGANDA (99.8%)

Neutral (Not Detected)

• "UN reports document civilian casualties in the conflict" → NO_PROPAGANDA (99.9%)
• "International observers documented attacks on hospitals" → NO_PROPAGANDA (99.9%)
• "The ICC is investigating potential war crimes" → NO_PROPAGANDA (99.9%)

---

For more information about the full propaganda detection system, visit the GitHub repository.