Headline Result
99.45%
F1 score · offline Linear SVM
Python
NLP
LLM Ensemble
Streamlit
Read on GitHub arrow_outward
Fake News Detection System
Problem
A single unverified claim spreads faster than the truth, and manual fact-checking can't keep pace with the volume of news and social posts.
Approach
A 7-source ensemble — Google Fact Check, Gemini 2.5, GPT-4o web search, NewsAPI, HuggingFace RoBERTa and a local TF-IDF + SVM — fused by weighted consensus voting. Accepts text, URLs, PDFs and even screenshots via GPT-4o Vision.
Result
The offline Linear SVM reaches 99.45% F1 on the ~44,000-article ISOT dataset (0.0055 overfit gap) and runs with no API key, while the live ensemble adds high-confidence, explainable verdicts.