|
|
---
|
|
|
language: multilingual
|
|
|
license: apache-2.0
|
|
|
tags:
|
|
|
- political-bias
|
|
|
- bias-detection
|
|
|
- news-analysis
|
|
|
- xlm-roberta
|
|
|
pipeline_tag: text-classification
|
|
|
---
|
|
|
|
|
|
# WorldVue Balanced Political Bias Detector
|
|
|
|
|
|
This model detects political bias in news articles across economic and social dimensions.
|
|
|
|
|
|
## Model Description
|
|
|
|
|
|
- **Architecture**: mDeBERTa-v3-base (278M parameters)
|
|
|
- **Training Data**: 6,000 articles judged by GPT-4o-mini
|
|
|
- **Languages**: 100+ (multilingual)
|
|
|
- **Axes**:
|
|
|
- Economic: LEFT (-1) ← CENTER (0) → RIGHT (+1)
|
|
|
- Social: LIBERTARIAN (-1) ← CENTER (0) → AUTHORITARIAN (+1)
|
|
|
|
|
|
## Signals Detected
|
|
|
|
|
|
**Economic (4 signals):**
|
|
|
1. Economic Role of State
|
|
|
2. Market & Business
|
|
|
3. Taxation & Spending
|
|
|
4. Labor & Trade
|
|
|
|
|
|
**Social (4 signals):**
|
|
|
1. State Authority vs Liberty
|
|
|
2. Cultural Identity
|
|
|
3. Immigration & Borders
|
|
|
4. Collective vs Individual
|
|
|
|
|
|
## Usage
|
|
|
|
|
|
```python
|
|
|
from transformers import AutoTokenizer, AutoModel
|
|
|
import torch
|
|
|
import torch.nn as nn
|
|
|
|
|
|
# Load model
|
|
|
tokenizer = AutoTokenizer.from_pretrained("Xiameineedsgpu/worldvue-balanced-bias-detector")
|
|
|
# ... (see example code below)
|
|
|
```
|
|
|
|
|
|
## Training Details
|
|
|
|
|
|
- **Test Loss**: 2.13
|
|
|
- **Political Articles**: 3,112
|
|
|
- **Non-Political Articles**: 2,888
|
|
|
- **Training Time**: ~1 hour on Google Colab (T4 GPU)
|
|
|
|
|
|
## Example Outputs
|
|
|
|
|
|
| Article | Economic | Social | Label |
|
|
|
|---------|----------|--------|-------|
|
|
|
| Fox News - Immigration | +0.02 | +0.83 | CENTER / AUTHORITARIAN |
|
|
|
| Energy Transition | -0.24 | -0.23 | LEFT / LIBERTARIAN |
|
|
|
| UK PPE Scandal | +0.10 | +0.25 | RIGHT / AUTHORITARIAN |
|
|
|
|
|
|
## License
|
|
|
|
|
|
Apache 2.0
|
|
|
|
