Skip to content

Evaluates and scores article credibility using a rules-based engine and fine-tuned transformer-based LLMs

Notifications You must be signed in to change notification settings

b00rg/article-credibility

Repository files navigation

Credibility Classifier

  • A credibility classifier built using a fine-tuned LLaMA model and Hugging Face Transformers, achieving 99% accuracy for labelling credible vs. non-credible articles.
  • Article Scores are accumulated for an overall publisher average credibility rating.
  • Created in collaboration with Quantexa as a foundation for an upcoming publisher credibility product.

Please note that as the model weights were trained on Quantexa datasets, these are unavailable for display. Feel free to train your own model weights using the finetuning folder code. See https://www.quantexa.com/ for more information.

Model Optimization

LoRA (Low-Rank Adaptation) and Mixed Precision Training reduced training time by 70% without impacting accuracy, allowing the model to be trained on free Kaggle GPUs in just 3 hours.

Publisher Scoring System:

  • A scoring mechanism is used to evaluate the overall credibility of news publishers based on historical content analysis in local storage.
  • Quick download/upload/clear buttons of publisher datasets for ease of use.

Interactive Frontend:

React-based dashboard to visualize credibility scores, trends, metrics, and publisher tables for the UI.

Reliable Article/Publisher

Screen.Recording.2025-04-08.at.12.34.01.mov

Unreliable Article/Publisher

Screen.Recording.2025-04-08.at.12.41.41.mov

About

Evaluates and scores article credibility using a rules-based engine and fine-tuned transformer-based LLMs

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published