This is a state-of-the-art fake news detection model based on DeBERTa-v3-large, achieving 99.98% accuracy on validation data. The model was fine-tuned on a carefully curated and deduplicated dataset combining multiple high-quality fake news datasets, totaling 51,319 samples after preprocessing.

🚀 Performance Highlights

Validation Accuracy: 99.98%
Test Accuracy: 99.94%
F1-Score: 99.98%
Precision: 99.97%
Recall: 100.00%

Model Architecture

Base Model: microsoft/deberta-v3-large
Task: Binary Text Classification (Real vs Fake News)
Parameters: ~400M parameters
Training Hardware: NVIDIA A100-SXM4-80GB

Training Details

Training Steps: 640
Batch Size: 64
Learning Rate: 3e-05
Max Length: 512 tokens
Training Time: 0.43 hours
Gradient Checkpointing: Non-reentrant (memory optimized)

Dataset Information

Total Samples: 51,319

Training: 41,055 samples
Validation: 5,132 samples
Test: 5,132 samples
Fake News: 30,123 samples
Real News: 21,196 samples Source Datasets:
mrisdal/fake-news
jainpooja/fake-news-detection
clmentbisaillon/fake-and-real-news-dataset

Usage

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
model_name = "Arko007/fact-check1-v1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Example prediction function
def predict_fake_news(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
    with torch.no_grad():
        outputs = model(**inputs)
        probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)
        prediction = torch.argmax(probabilities, dim=-1).item()

    labels = {0: "REAL", 1: "FAKE"}
    confidence = probabilities[0][prediction].item()

    return {
        "prediction": labels[prediction],
        "confidence": confidence,
        "probabilities": {
            "REAL": probabilities[0][0].item(),
            "FAKE": probabilities[0][1].item()
        }
    }

# Test the model
text = "Breaking: Scientists discover new planet in our solar system!"
result = predict_fake_news(text)
print(f"Prediction: {result['prediction']} ({result['confidence']:.2%} confidence)")

Model Performance

This model achieves research-grade performance on fake news detection, with near-perfect accuracy across all metrics. The high precision and recall indicate excellent balance between catching fake news while avoiding false positives on real news.

Limitations and Bias

Trained primarily on English news articles
Performance may vary on news domains not represented in training data
May reflect biases present in the source datasets
Designed for binary classification (fake vs real) only

Citation

bibtex

@misc{fake-news-deberta-2025,
author = {Arko007},
title = {Elite Fake News Detection with DeBERTa-v3-Large},
year = {2025},
publisher = {Hugging Face},
url = {[https://huggingface.co/](https://huggingface.co/)Arko007/fact-check1-v1}
}

License

MIT License - Feel free to use this model for research and applications.

Built with ❤️ using A100 80GB + DeBERTa-v3-Large