abullard1/germeval-2025-harmful-content-detection-training-dataset

GermEval 2025 Harmful Content Detection - Training Sets

(Call to Action • Attacks on Democratic Basic Order • Violence)

Author: Samuel Ruairí Bullard - University of Regensburg

Models: Model Zoo (Gradio Space)

Base model: LSX-UniWue/ModernGBERT_134M

Competition: GermEval 2025 Shared Task

Collection: GermEval 2025 Contribution Collection

abullardUR@GermEval Shared Task 2025 Submission

---

Dataset Summary

This repository republishes the training splits used in the GermEval 2025 Shared Task on harmful content detection in German social media (Twitter). The three subtasks are:

• Call to Action (C2A): Binary detection of risky calls to action
• Attacks on the Democratic Basic Order (DBO): Four-class classification (nothing, criticism, agitation, subversive)
• Violence (VIO): Binary detection of violence-related content

Data consists of German tweets from right-wing extremist networks (2014–2016), annotated by domain experts. The task emphasizes handling severe class imbalance and primarily short, noisy Tweets.

Files Included

• c2a_train.csv (~9,084 rows; header, comma-separated)
• dbo_train.csv (~9,617 rows; header, comma-separated)
• vio_train.csv (~10,230 rows; header, comma-separated)

All CSVs are comma-separated with the following schema:

• id: tweet identifier (string)
• description: tweet text (string)
• C2A / DBO / VIO: task-specific label

Label values:

• C2A: "TRUE" or "FALSE"
• DBO: "nothing", "criticism", "agitation", "subversive"
• VIO: "TRUE" or "FALSE"

How to Load

You can load directly using the Hugging Face datasets library:

from datasets import load_dataset

# Option A: Load a specific config's train split via CSV builder
c2a = load_dataset(
    "csv",
    data_files={"train": "hf://datasets/abullard1/germeval-2025-harmful-content-detection-training-dataset/c2a_train.csv"}
)["train"]

dbo = load_dataset(
    "csv",
    data_files={"train": "hf://datasets/abullard1/germeval-2025-harmful-content-detection-training-dataset/dbo_train.csv"}
)["train"]

vio = load_dataset(
    "csv",
    data_files={"train": "hf://datasets/abullard1/germeval-2025-harmful-content-detection-training-dataset/vio_train.csv"}
)["train"]

# Option B: Pandas
import pandas as pd
c2a_df = pd.read_csv("https://huggingface.co/datasets/abullard1/germeval-2025-harmful-content-detection-training-dataset/resolve/main/c2a_train.csv")

Ethical and Safety Considerations

⚠️ WARNING: This dataset contains harmful, extremist, violent, and offensive language. It is intended strictly for research on harmful content detection. Exercise caution when displaying examples in user-facing interfaces!

Licensing & Usage

This dataset is licensed under GPL-3.0 and requires attribution to both the GermEval 2025 organizers (University of Stuttgart, University of Mannheim) and data collectors (Mittweida University of Applied Sciences) when used.

Suggested Citation

Please cite the GermEval 2025 Shared Task and this dataset:

@inproceedings{germeval2025sharedtask,
  title={GermEval 2025 Shared Task: Harmful Content Detection on German Social Media},
  year={2025},
  note={Competition dataset; see competition website for details},
  url={https://www.codabench.org/competitions/4963/}
}

@misc{abullard2025germevaltraining,
  author={Samuel Ruairí Bullard},
  title={GermEval 2025 Harmful Content Detection - Training Splits},
  year={2025},
  howpublished={Hugging Face Dataset},
  url={https://huggingface.co/datasets/abullard1/germeval-2025-harmful-content-detection-training-dataset}
}