AI-Fake_News_Detection_Roberta

Fake News Detection Model

A RoBERTa-based text classification model for detecting fake news articles. This model was evaluated on the WELFake dataset as part of a final year project focused on misinformation detection using NLP.

Model Details

Model Description

• Developed by: Akan Brown
• Model type: RoBERTa (Robustly Optimized BERT Pretraining Approach)
• Language: English
• License: MIT
• Base model: jy46604790/Fake-News-Bert-Detect
• Fine-tuned from: FacebookAI/roberta-large

Model Sources

• Base Model Repository: https://huggingface.co/jy46604790/Fake-News-Bert-Detect
• Dataset: https://www.kaggle.com/datasets/saurabhshahane/fake-news-classification

---

Uses

Direct Use

This model can be used directly for binary classification of news articles as either FAKE or REAL without any additional fine-tuning.

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="Akan-Brown/fakenews-detector",
    truncation=True,
    max_length=512
)

result = classifier("Your news article text here")[0]
label = "REAL" if result["label"] == "LABEL_1" else "FAKE"
print(f"Prediction: {label} ({result['score']:.2%} confidence)")

Downstream Use

• Fake news detection APIs
• Browser extensions for real-time news verification
• Media literacy tools and platforms
• Academic research on misinformation

Out-of-Scope Use

• Non-English news articles
• Satire detection (satirical news may be misclassified as fake)
• Real-time social media posts (model was trained on article-length text)
• Opinion pieces and editorials

---

Label Mapping

---

Bias, Risks, and Limitations

• The model was trained primarily on English-language news data and may not generalise well to other languages or dialects.
• The WELFake dataset has a bias toward US and UK political news — performance may vary on news from other regions.
• The model may struggle with satire, opinion pieces, or highly technical journalism.
• As with all ML models, this should not be used as the sole arbiter of whether a news article is fake or real.

Recommendations

This model should be used as a supplementary tool alongside human judgment and other fact-checking resources. It is not suitable for automated content moderation without human oversight.

---

Training and Evaluation Details

Dataset

The model was evaluated on the WELFake dataset, a benchmark dataset for fake news detection containing 72,134 news articles:

• Real news: sourced from Reuters, The New York Times, The Washington Post, and Guardian
• Fake news: sourced from various unreliable sources flagged by fact-checkers
• Split used: 80% train / 10% validation / 10% test

Evaluation Results on WELFake (1,000 sample test)

Replace the values above with your actual results from Cell 7 output.

---

How to Get Started

Basic prediction

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="Akan-Brown/fakenews-detector",
    truncation=True,
    max_length=512
)

articles = [
    "Federal Reserve raises interest rates by 0.25% amid inflation concerns.",
    "SHOCKING: Government putting microchips in vaccines, whistleblower reveals!!!",
]

for article in articles:
    result = classifier(article)[0]
    label = "REAL ✅" if result["label"] == "LABEL_1" else "FAKE ❌"
    print(f"{label} ({result['score']:.2%}) — {article[:60]}")

Using the Inference API

import os
from huggingface_hub import InferenceClient

client = InferenceClient(
    provider="hf-inference",
    api_key=os.environ["HF_TOKEN"],
)

result = client.text_classification(
    "Your news article text here",
    model="Akan-Brown/fakenews-detector",
)

---

Project Context

This model was developed as part of a Final Year Project on automated fake news detection using transformer-based NLP models. The project explores the use of pre-trained language models for identifying misinformation in online news articles.

The full system includes:

• A fine-tuned RoBERTa classification model (this repo)
• A FastAPI backend deployed on Render
• A web frontend for real-time news verification

---

Citation

If you use this model in your research, please cite the WELFake dataset:

@article{verma2021WELFake,
  title={WELFake: Word Embedding over Linguistic Features for Fake News Detection},
  author={Verma, Parth and Agrawal, Priya},
  journal={IEEE Transactions on Computational Social Systems},
  year={2021}
}

---

Model Card Contact

Akan Brown — Final Year Project, 2026