This project trains a BERT-based classifier to detect fake news using HuggingFace Transformers and PyTorch.

• Fake.csv – Dataset of fake news articles
• True.csv – Dataset of real news articles
• main.py – Main training and evaluation script
• data_analysis.ipynb – Notebook for dataset exploration and visualization
• bert-base-uncased/ – (Optional) Local BERT model directory (or use HuggingFace download)

Make sure you have Python 3.7+ and install the following packages:

pip install transformers datasets scikit-learn pandas numpy torch nltk

You also need to download NLTK stopwords:

import nltk
nltk.download('stopwords')

Run data_analysis.ipynb to explore and visualize the dataset. It performs the following:

• Loads and merges Fake.csv and True.csv
• Assigns labels (0 = Fake, 1 = Real)
• Samples 3000 articles for faster experimentation
• Visualizes class distribution:
  • Fake vs Real label balance (relatively balanced)
  • Distribution of subject categories by label (unbalanced, not used as feature)
• Observes that the date field contains some noisy or invalid strings (e.g., URLs), so date is excluded as a feature

The analysis helps confirm that only the text field is suitable as a classification input.

1. Make sure Fake.csv and True.csv are in the same folder as main.py.
2. (Optional) If using a local model, ensure bert-base-uncased/ is in the same directory and modify this line in the code:

bert_name = "./bert-base-uncased" # path to your local model 

​ Otherwise, the model will be downloaded automatically from HuggingFace.

3. Run training:

python main.py

After training, the best model is saved in the ./results directory and evaluation metrics will be printed.

• Training progress and validation metrics printed during training
• Final test accuracy, precision, recall, and F1 score