README.md

---
license: cc-by-4.0
language:
- de
base_model:
- dbmdz/bert-base-german-cased
pipeline_tag: text-classification
---

# SIP-BERT

**SIP-BERT** is a transformer-based model designed to detect **social inequality** in German texts.  
It was fine-tuned on **German Bundestag debates** (sourced from [OpenDiscourse](https://doi.org/10.7910/DVN/FIKIBO)), where each training instance consists of 3-sentence segments.

---

## Model Description

- **Architecture**: `bert-base-german-cased` (from [dbmdz](https://huggingface.co/dbmdz/bert-base-german-cased))  
- **Task**: Binary classification – detecting social inequality in German texts  
- **Labels**:  
  - `0` = no social inequality  
  - `1` = social inequality  
- **Language**: German
- **Training Data**: 1,950 annotated text passages from Bundestag debates (via OpenDiscourse)  
- **Segmenting**: Data split into 3-sentence units
- **Evaluation**: Accuracy 0.97; F1 Score 0.95 

---

## Intended Use

- **Primary use case**: Analysis of parliamentary discourse on social inequality  
- **Research contexts**: Political science, computational social science, discourse analysis  

---

## Limitations

- The model is trained on Bundestag debates (1949–2021), but is **specialized for texts from 1990 onwards**.  
- It may be less reliable for earlier parliamentary language (1949–1989) and for **non-parliamentary speech**.  
- It was designed primarily to detect **economic inequality**, and it may not be applicable to other types of inequality.

---

## Usage

You can load the model with the Hugging Face `transformers` library:

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("miriamex/SIP-BERT")
model = AutoModelForSequenceClassification.from_pretrained("miriamex/SIP-BERT")

inputs = tokenizer("Hier ein Beispieltext über soziale Ungleichheit.", return_tensors="pt")
outputs = model(**inputs)
```