Sumra81/fake-news-distilbert

Fake News Detector (DistilBERT)

A DistilBERT model fine-tuned to tell whether a news article's writing style looks real or fake.

Labels

• 0 = Fake
• 1 = Real

Training data

• ISOT Fake and Real News dataset (15,000 articles used)
• WELFake dataset (10,000 articles used, labels flipped so that 0 = fake, 1 = real)
• Text was cleaned first: lowercased, links removed, and agency tags, dates, month and day names, and city names removed to reduce shortcut clues.

Training setup

• Base model: distilbert-base-uncased
• 2 epochs, batch size 16, learning rate 2e-5, max length 256 tokens
• Trained on a free Kaggle GPU

Results

Test	Accuracy	F1
Held-out WELFake (trained on ISOT + WELFake)	96.57%	0.964
WELFake, when trained on ISOT only	86.00%	0.842

The 86% figure is the better sign of how it handles a source it never trained on.

Limits

• It reads writing style, not facts. A false story written like a news-agency report can be marked real.
• Accuracy drops sharply on sources that look different from the training data.
• Only about the first 200 words of an article are read.
• Not for real fact-checking or moderation decisions.

How to use

from transformers import pipeline
clf = pipeline("text-classification", model="Sumra81/fake-news-distilbert")
print(clf("Paste a cleaned article here"))