loading
loading
The measured fraction of outputs where the model states something false or unsupported.
You turn a fuzzy worry ('it makes things up') into a tracked metric by scoring a sample of outputs — via human review or an LLM judge checking each claim against the source — and reporting the percentage that are fabricated or contradicted by context. It matters because you can't improve what you don't measure, and it lets you compare prompts, models, and RAG setups objectively. Example: in a RAG bot, 6% of answers cite facts not present in the retrieved documents.
Plainly
Think of Hallucination Rate as the checklist that keeps the app open for real people. The measured fraction of outputs where the model states something false or unsupported.
In practice
Use it when a change has to survive deploys, users, incidents, analytics, or billing reality. In practice, define the owner, input, output, and failure mode before you rely on it.