library_name: transformers
datasets:
- vector-institute/newsmediabias-plus
language:
- en
metrics:
- accuracy
- precision
- recall
- f1
base_model:
- google-bert/bert-base-uncased
pipeline_tag: text-classification
BERT NMB+ (Disinformation Sequence Classification):
Classifies 512 chunks of a news article as "Likely" or "Unlikely" biased/disinformation.
Fine-tuned BERT (bert-base-uncased) on the headline, aritcle_text and text_label fields in the News Media Bias Plus Dataset.
This model was trained without weighted sampling, and the dataset contains 81.9% 'Likely' and 18.1% 'Unlikely' examples. The same model trained with weighted sampling preformed worse on training eval metrics, but better when evaluated by gpt-4o-mini as a judge and is available here.
Metics
Evaluated on a 0.1 random sample of the NMB+ dataset, unseen during training
- Accuracy: 0.7884
- Precision: 0.8573
- Recall: 0.8599
- F1 Score: 0.8586
How to Use:
Keep in mind, this model was trained on full 512 token chunks (tends to over-predict Unlikely for standalone sentences). If you're planning on processing stand alone sentences, you may find better results with this NMB+ model, which was trained on biased headlines.
from transformers import pipeline
classifier = pipeline("text-classification", model="maximuspowers/nmbp-bert-full-articles")
result = classifier("He was a terrible politician.", top_k=2)
Example Response:
[
{
'label': 'Likely',
'score': 0.9967995882034302
},
{
'label': 'Unlikely',
'score': 0.003200419945642352
}
]