sciGNAD_tcls

ChristBERT/sciGNAD

Model Description

ChristBERT/sciGNAD is a German binary text classification model for distinguishing scientific content from general-domain text.

The model is based on GeistBERT base and fine-tuned for domain filtering tasks, enabling reliable separation of relevant scientific content from non-scientific text.

---

Intended Use

Primary Use Cases

• Classification of German text into:
  • scientific/medical
  • non-scientific
• Data filtering for domain-specific NLP pipelines
• Preselection of relevant documents in retrieval systems

Out-of-Scope Use

• Fine-grained topic classification
• Clinical or medical decision support
• High-stakes applications

---

Training Data

The model was fine-tuned on a binary-labeled dataset derived from the 10kGNAD.

• Total dataset size: 10,000 German news articles
• Scientific subset: 573 articles

Dataset Construction

• Positive class:
  • Scientific articles from 10kGNAD
• Negative class:
  • Stratified sample from remaining categories
• Balanced dataset:
  • Equal class distribution

---

Evaluation

The model was evaluated on a manually labeled dataset:

• Test size: 119 documents
• Annotation tool: LabelStudio

Performance

---

Training Details

• Base model: GeistBERT_base
• Task: Binary classification
• Language: German
• Framework: Hugging Face Transformers

---

Limitations

• The model was trained on a dataset derived from German news articles, which may limit transferability to other text genres.
• The evaluation set was relatively small, consisting of 119 manually labeled web-crawl documents (medical field).
• The model is intended for filtering scientific/medical relevance and should not be used for clinical or medical decision-making.

---

Ethical Considerations

• Intended for filtering and preprocessing, not decision-making
• Misclassifications may affect downstream dataset quality

---

Citation

If you use this model, please cite:

@misc{christbert_scignad,
  title={ChristBERT/sciGNAD: Scientific Content Classifier for German Web Data},
  author={Schmitt, Raphael},
  year={2026}
}

@article{he_word_2026,
  author         = {He, Henry and Frei, Johann and Schmitt, Raphael},
  title          = {The Word and the Way: Strategies for Domain-Specific {BERT} Pre-Training in {German} Medical {NLP}},
  journal        = {BMC Medical Informatics and Decision Making},
  year           = {2026},
  volume         = {26},
  article-number = {296},
  doi            = {10.1186/s12911-026-03728-2},
  url            = {https://doi.org/10.1186/s12911-026-03728-2},
  issn           = {1472-6947}
}