Skip to content

Dataset · PatronusAI

Hallucination benchmark for evaluating language models

PatronusAI/HaluBench

15,000 labeled English context-question-answer triplets for evaluating language models on hallucination tendencies.

What's inside

The dataset consists of 15,000 English text examples in a context-question-answer format, where each case indicates whether the response contains a hallucination. The examples are gathered from existing QA datasets and cover real-world domains, including finance and medicine.

What you can do with it

  • Evaluating and comparing LLMs for hallucination resistance in critical domains
  • Training detector models to filter out unreliable responses in RAG systems
Tasks
Text generation
Languages
English
Size
10K–100K rows
Downloads this month
1.2K
Likes
46

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("PatronusAI/HaluBench")