abullard1/germeval2025-dbo-moderngbert-cw_and_focal

🏆 GermEval 2025: Democratic Basic Order Attack Detection (Class-Weighted + Focal Loss)

abullardUR@GermEval Shared Task 2025 Submission

---

🎯 Model Summary

This model is a fine-tuned version of LSX-UniWue/ModernGBERT_134M, specifically designed for Attacks on Democratic Basic Order (DBO) Detection in German social media content. It was developed as part of the GermEval 2025 Shared Task on Harmful Content Detection.

🏅 Competition Performance

• Final Ranking: 6th out of 9 teams
• Primary Metric (Macro-F1): 0.56 (+19% over official baseline (0.47 → 0.56))
• Approach: Combined class-weighted cross-entropy + focal loss for extreme multi-class imbalance

📊 Task Details

• Task Type: Multi-class classification (4 classes)
• Classes:
  • Nothing (84.2%): No attack on democratic order
  • Criticism (10.8%): Legitimate criticism
  • Agitation (4.2%): Harmful agitation
  • Subversive (0.8%): Most severe attacks on democratic principles
• Domain: German social media (Twitter, 2014-2016)
• Data Source: Right-wing extremist network posts

⚠️ Limitations and Bias

Known Limitations

• Extreme Class Imbalance: Subversive class represents only 0.8% of data (60 samples)
• Focal Loss Limitations: May inadvertently down-weight important minority examples
• Domain Specificity: Trained on 2014-2016 German Twitter data from right-wing extremist networks
• Temporal Bias: Language patterns may not reflect contemporary usage

Ethical Considerations

• Model trained on potentially harmful content for research purposes only
• Should not be used to amplify or generate harmful content
• Requires careful handling due to sensitive training data

🚀 How to Use

Quick Start

from transformers import AutoProcessor, AutoModelForSequenceClassification

# Load model and processor
model_id = "abullard1/germeval2025-dbo-moderngbert-cw_and_focal"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForSequenceClassification.from_pretrained(model_id, trust_remote_code=True).eval()

# Run inference
text = "Die aktuelle Regierung sollte ihre Politik überdenken."
inputs = processor(text, return_tensors="pt", truncation=True)
probs = model(**inputs).logits.softmax(-1).detach().cpu().numpy()
print(f"Predictions: {probs}")

Class Labels

• Label 0: Nothing (no attack on democratic order)
• Label 1: Criticism (legitimate criticism)
• Label 2: Agitation (harmful agitation against democratic order)
• Label 3: Subversive (severe attacks on democratic principles)

📈 Training Details

Training Data

• Source: GermEval 2025 Shared Task DBO dataset
• Size: 7,454 samples
• Split: 80% training (5,963), 20% validation (1,491)
• Class Distribution:
  • Nothing: 6,277 (84.2%)
  • Criticism: 804 (10.8%)
  • Agitation: 313 (4.2%)
  • Subversive: 60 (0.8%)

Training Procedure

• Base Model: ModernGBERT-134M (8192 token context)
• Architecture: Mean-pooling classification head
• Loss Function: Class-weighted cross-entropy + Focal Loss
• Optimizer: AdamW with linear scheduling
• Early Stopping: Patience of 5 epochs on validation Macro-F1

Hyperparameters

• Learning Rate: 3e-5
• Weight Decay: 0.0270
• Batch Size: 16/16 (train/eval)
• Epochs: 8
• Warmup Steps: 100
• Focal Loss Gamma: 1.537

📚 Citation

@inproceedings{bullard2025germeval,
  title   = {abullardUR@GermEval Shared Task 2025: Fine-tuning ModernGBERT on Highly Imbalanced German Social Media for Harmful Content Detection},
  author  = {Bullard, Samuel},
  year    = {2025},
  booktitle = {Proceedings of KONVENS 2025 Workshops}
}

🙏 Acknowledgments

• GermEval 2025 Organizers: University of Stuttgart and University of Mannheim
• Prof. Dr. Udo Kruschwitz (University of Regensburg) for supervision
• ModernGBERT Team: LSX-UniWue for the ModernGBERT-134M German language base-model

📄 License

This model inherits the Research-only RAIL-M license from ModernGBERT. See license details.