RealFakeNews: A Dataset for Detecting Fake News
RealFakeNews is a dataset of over 108,000 news samples, created to support the development of models that detect misinformation. Each entry contains a short news article along with a label indicating whether it’s real or fake.
What's in the Dataset?
- Samples: 108,032
- Columns:
text: News content (string)label: Classification label (string:REALorFAKE)
- Language: English
- Format: CSV
- License: CC BY‑NC‑SA 4.0
Label Distribution
| Label | Meaning | Count |
|---|---|---|
| REAL | Real News | 64,641 |
| FAKE | Fake News | 43,391 |
The dataset is slightly imbalanced, with more real news than fake news entries.
Use Cases
- Fake news classification
- NLP experiments on misinformation
- Training and fine‑tuning transformers (e.g., BERT, RoBERTa)
- Evaluation using classification metrics like Accuracy, F1-score, ROC-AUC
Sample Entry
{
"text": "From India Censoring Internet Archive To No Night On Aug 12: Not Real....No! Banks are NOT charging Rs 150 after 4 transactions on...",
"label": "FAKE"
}
⚠️ Warning
- Misinformation Risk: This dataset includes content labeled as "FAKE" that may contain false or misleading information. Even "REAL" entries may not be fully accurate or current.
- Not for Automated Fact-Checking: Models trained using this dataset should not be used to make critical decisions or perform real-world fact-checking without human oversight.
- Potential Bias: While the data comes from a range of sources, biases may still exist in topic selection, language, or cultural framing.
- Research & Educational Use Only: This dataset is intended strictly for non-commercial research and educational purposes. Commercial use requires separate permission.
Use this dataset responsibly and with awareness of its limitations.