abullard1/abullardUR_GermEval2025_Submission_ModelZoo

🏆 GermEval 2025 Harmful Content Detection Model Zoo

Interactive Gradio Space for testing all 6 models submitted to the GermEval 2025 Shared Task on Harmful Content Detection.

🎯 Features

• —6 Pre-trained Models: Test all submitted models (Class-Weighted vs Focal Loss approaches)
• —3 Detection Tasks: Call to Action, Democratic Basic Order Attacks, Violence Detection

🚀 Models Available

📢 Call to Action Detection (C2A)

• —Class-Weighted: Macro-F1 0.82
• —Class-Weighted + Focal Loss: Macro-F1 0.82

🏛️ Democratic Basic Order Attacks (DBO)

• —Class-Weighted: Macro-F1 0.63
• —Class-Weighted + Focal Loss: Macro-F1 0.56

⚔️ Violence Detection (VIO)

• —Class-Weighted: Macro-F1 0.82
• —Class-Weighted + Focal Loss: Macro-F1 0.81

🔬 Research Context

This model zoo showcases our submission to the GermEval 2025 Shared Task, focusing on detecting harmful content in German social media. The models are fine-tuned versions of ModernGBERT-134M, specifically adapted for the challenges of:

• —Extreme class imbalance
• —German language nuances in social media
• —Subtle harmful content detection

Resources

All resources (Code, Data, Evaluation) concerning our GermEval 2025 contribution can be found in the GitHub repository 🐱.

📚 Citation

If you use these models in your research, please cite our GermEval 2025 submission paper (link coming soon).

Author: Samuel Ruairí Bullard (@abullardUR) Institution: University of Regensburg Supervisor: Prof. Dr. Udo Kruschwitz