RefChecker – Hallucination Detection

Fine-grained detection of factual errors in large language model outputs.

  • Duration: August 2023 – November 2023
  • Technologies: Python, PyTorch, NLP, LLMs

View code on GitHub

Highlights

  • Engineered an automated hallucination-detection system that identifies factual errors in LLM outputs at the sub-sentence level using contextual entailment.
  • Achieved 90% precision on benchmark datasets with a modular architecture combining extractor, checker, and aggregator components.