HallucinationBenchmark<T>

Models & Types · AiDotNet.Safety.Benchmarking

Benchmark for evaluating hallucination and factual grounding detection accuracy.

Inheritance: SafetyBenchmarkBase<T> → HallucinationBenchmark<T>

For Beginners

“Hallucination” is when an AI makes something up that sounds real but isn’t — like inventing a fake research paper or citing false statistics. This benchmark tests whether your safety system can detect these fabrications.

How It Works

Tests the safety pipeline against text containing factual claims, fabricated citations, invented statistics, and self-contradictions. Also includes factually grounded text to measure false positive rates.

References:

  • HaluEval: Hallucination evaluation for LLMs (ACL 2023)
  • FActScore: Fine-grained atomic evaluation of factual precision (EMNLP 2023)
  • FELM: Benchmarking factuality evaluation of LLMs (NeurIPS 2024)

Constructors

Constructor Summary
HallucinationBenchmark()

Properties

Property Summary
string BenchmarkName { get; }

Methods

Method Summary
SafetyBenchmarkResult RunBenchmark(SafetyPipeline<T> pipeline)