📰 Fake News Detection with Fine-Tuned DistilBERT
This repository contains a fine-tuned version of DistilBERT for fake news detection,
trained on the LIAR dataset. The model predicts credibility scores on a scale of 0 to 5, rather than just a binary classification.
This model is a fine-tuned version of
distilbert-base-uncased on
the fake.csv & true.csv (Binary classification) dataset.
It achieves the following results on the evaluation set:
- Loss: 0.0001
- Accuracy: 1.0
Model Overview
Model: Fine-tuned DistilBERT (distilbert-base-uncased)
Dataset: fake.csv & true.csv (Binary classification)
Labels:
0 → Real News 📰
1 → Fake News 🚨
Use Case: Detect misinformation and classify news as real or fake
Deployment: Hosted on Hugging Face and accessible via FastAPI
Intended uses & limitations
✅ Intended Uses
Fake News Detection: Classifies news articles as FAKE (1) or REAL (0).
Misinformation Analysis: Helps identify misleading or fabricated news.
News Verification Tools: Can be integrated into fact-checking websites, browser extensions, or news aggregator platforms.
Educational Purposes: Useful for research and academic projects on NLP and misinformation detection.
More information needed
⚠️ Limitations
Limited Dataset: Trained only on fake.csv and true.csv, which may not represent all news sources.
Context Understanding: The model classifies based on text patterns, not fact-checking against external sources.
Adversarial Attacks: It may be fooled by well-crafted fake news.
Bias & Generalization: Might struggle with different writing styles, languages, or new topics outside its training data.
Continuous Updates Needed: Fake news trends evolve, requiring periodic retraining with fresh data.
Training and evaluation data
The model was trained on a binary classification dataset consisting of:
fake.csv → Contains fake news articles
true.csv → Contains real news articles
🔹 Data Preprocessing
Text Cleaning: Removed special characters, extra spaces, and stopwords.
Tokenization: Used DistilBERT tokenizer to convert text into input tokens.
Label Encoding: Assigned 1 for fake news and 0 for real news.
Train-Test Split: 80% for training, 20% for evaluation.
📈 Evaluation Metrics
The model was evaluated using:
Accuracy
Precision, Recall, F1-score
Loss (Cross-Entropy Loss)
Training procedure
🔹 Training Process
Loaded fake.csv and true.csv → Preprocessed the text
Used Hugging Face’s Trainer API to fine-tune DistilBERT
Monitored validation loss & accuracy after each epoch
Saved the best-performing model to Hugging Face
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 3e-05
- train_batch_size: 32
- eval_batch_size: 64
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 64
- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- num_epochs: 5
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Accuracy |
|---|
| 0.0021 | 1.0 | 395 | 0.0089 | 0.9983 |
| 0.0034 | 2.0 | 790 | 0.0008 | 0.9997 |
| 0.0001 | 3.0 | 1185 | 0.0000 | 1.0 |
| 0.0 | 4.0 | 1580 | 0.0006 | 0.9997 |
| 0.0 | 5.0 | 1975 | 0.0001 | 1.0 |
🚀 API Deployment with FastAPI
The model is already deployed using FastAPI for real-time predictions.
🔗 API Endpoint
🛠 How to Use the API
You can send a POST request to the API to classify news as FAKE (1) or REAL (0).
Example Request:
curl -X POST "
https://anup069-fake-news-detection-api.hf.space/verify/"
-H "Content-Type: application/json"
-d '{"text": "Breaking news: Scientists discover a new planet!"}'
Example Response:
{
"label": "FAKE",
"score": 0.98
}
✅ Features
✔️ Real-time inference with low latency
✔️ Scalable & FastAPI-based backend
✔️ Can be integrated into web apps, Chrome extensions, or other platforms
Framework versions
- Transformers 4.48.2
- Pytorch 2.5.1+cu124
- Datasets 3.2.0
- Tokenizers 0.21.0
