TruthEval: A Dataset to Evaluate LLM Truthfulness and Reliability

TL;DR

TruthEval is a carefully curated dataset designed to evaluate the truthfulness and reliability of Large Language Models across sensitive topics, revealing their limitations in understanding simple questions.

Contribution

This paper introduces TruthEval, a new benchmark dataset with challenging, truth-annotated statements to better assess LLMs' factual accuracy and reliability.

Findings

LLMs often fail in simple truth verification tasks

The dataset reveals limitations in LLM understanding of straightforward questions

Initial analyses show significant room for improvement in LLM truthfulness

Abstract

Large Language Model (LLM) evaluation is currently one of the most important areas of research, with existing benchmarks proving to be insufficient and not completely representative of LLMs' various capabilities. We present a curated collection of challenging statements on sensitive topics for LLM benchmarking called TruthEval. These statements were curated by hand and contain known truth values. The categories were chosen to distinguish LLMs' abilities from their stochastic nature. We perform some initial analyses using this dataset and find several instances of LLMs failing in simple tasks showing their inability to understand simple questions.