vector-institute/Unbias-plus

---
language:
- en
task_categories:
- text-classification
- token-classification
- other
task_ids:
- multi-label-classification
- named-entity-recognition
tags:
- bias-detection
- news
- disinformation
- debiasing
- responsible-ai
pretty_name: UnBias Plus Dataset
size_categories:
- 1K<n<10K
---

# UnBias Plus Dataset

## Dataset Summary

UnBias Plus is a dataset for news bias detection. It is built on top of [VLDBench](https://huggingface.co/datasets/vector-institute/VLDBench), using the same source articles, extended with bias-specific annotations including severity scores, span-level biased segment highlights, and debiased rewrites.

## Dataset Structure

| Column | Type | Description |
|---|---|---|
| index | int64 | Row index |
| binary_label | string | Bias label: `biased` or `unbiased` |
| article_text | string | Full article content |
| severity | int64 | Bias severity score (0–4) |
| bias_found | bool | Whether bias was detected |
| biased_segments | list | Span-level biased text annotations |
| unbiased_text | string | Debiased rewrite of the article |

**Split:** train · 4,867 rows

## Usage

```python
from datasets import load_dataset

ds = load_dataset("vector-institute/Unbias-plus", split="train")

record = ds[0]
print(record['binary_label'])
print(record['severity'])
print(record['biased_segments'])
```

This dataset can be used for:
- News bias detection and classification
- Bias severity estimation
- Span-level bias annotation (token classification)
- Debiasing (seq2seq rewriting)

## Source Data

Articles are sourced from the same news outlets as VLDBench. See the [VLDBench dataset](https://huggingface.co/datasets/vector-institute/VLDBench) and accompanying [paper](https://arxiv.org/abs/2502.11361) for full details on data collection.

## Contact

For inquiries, contact Shaina Raza, PhD at [e-mail]