What AI says about Hallucination Benchmark Suite

This repository provides a notebook-driven workflow to build ground-truth QA data, inject controlled hallucinations, and benchmark hallucination-detection frameworks. It builds a QA dataset from RAGBench-COVIDQA, generates hallucinated examples, and runs multiple detectors (OpenAI, Groq, UpTrain, TruthLLM) to evaluate them. It computes metrics like precision, recall, F1, ROC-AUC, and cost, and outputs plots and a Markdown leaderboard for reproducible comparisons across single models or entire toolchains, with the detection framework verifying facts against context using LLM evaluators.