ThunderstormBERT-de-v1

Model Card for ThunderBERT-de-v1

Model Summary

ThunderBERT-de-v1 is a BERT-based language model for multi-class classification of German-language texts that describe thunderstorm events.
Its purpose is to automatically detect and differentiate between varying intensities of thunderstorm-related content in historical and contemporary German sources.
The model was developed and trained during Phase B of the Thunderstorm Project at the University of Freiburg.

---

Model Details

Model Description

• Model name: ThunderBERT-de-v1
• Developed by: Franck Schätz, University of Freiburg
• Model type: Encoder-based transformer for multi-class sequence classification
• Language(s): German (de)
• License: Apache-2.0
• Fine-tuned from: bert-base-german-cased
• Number of labels: 4
• Labels:
  • GW0: no thunderstorm
  • GW1: light thunderstorm
  • GW2: strong thunderstorm
  • GW3: severe thunderstorm
• Intended granularity: Document- or passage-level classification

The model uses a SentencePiece-based tokenizer configuration (spm.model, tokenizer_config.json, etc.).
For compatibility with the training setup, it may be necessary to load the tokenizer with use_fast=False.

Model Sources

• Repository: https://huggingface.co/<namespace>/ThunderBERT-de-v1
  (replace <namespace> with your actual HF username or organization)
• Paper: (to be added if/when available)
• Demo: (optional, e.g. Spaces link if created)

---

Uses

Direct Use

This model is designed for automatic classification of thunderstorm intensities in German-language texts.
It assigns each input text (sentence, passage, or short paragraph) to one of four categories:

Typical applications include:

• Identifying and categorizing thunderstorm mentions in:
  • historical weather chronicles,
  • meteorological observations,
  • diaries, letters, or reports describing severe weather.
• Supporting large-scale extraction of climatological events for environmental or historical studies.

Downstream Use

Can be integrated into:

• event extraction or climate knowledge graph pipelines,
• cross-referencing with temporal and geographic entities (e.g. for spatio-temporal storm mapping),
• semi-automated annotation or corpus enrichment.

Out-of-Scope Use

Not intended for:

• Operational forecasting or nowcasting of thunderstorms,
• Real-time meteorological decision-making,
• Safety-critical or legal decision contexts,
• General-purpose sentiment or topic classification.

---

Bias, Risks, and Limitations

Limitations

• Domain specificity: Performance is optimized for historical and descriptive sources.
  Results may degrade for short, modern, informal, or technical texts.
• Language coverage: Only German is supported.
• Label granularity: The model captures thunderstorm intensity levels but not duration, geographic location, or damage.
• Data transparency: The dataset is internal to the project and was annotated by domain experts.

Potential Biases

• Possible overrepresentation of specific regions, historical periods, or linguistic styles.
• Ambiguities in wording (e.g., metaphorical "storm") can cause misclassification.
• OCR errors or historical orthography can reduce recall.

Recommendations

• Use as a research and discovery tool, not as an operational detector.
• Combine automated classification with manual validation or expert review.
• For other domains, additional fine-tuning is recommended.

---

How to Get Started with the Model

from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_id = "<namespace>/ThunderBERT-de-v1"  # e.g. "Stickmu/ThunderBERT-de-v1"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

text = "Am Abend entluden sich schwere Gewitter über dem Schwarzwald."
inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)

predicted_class_id = int(outputs.logits.argmax(dim=-1))
print(model.config.id2label[predicted_class_id])