hallucination-leaderboard

Why we included this project

If you are picking a model for summarization-heavy RAG or other document-grounded work, hallucination rate is often the number that matters more than a flashy benchmark score. Vectara feeds each LLM the same short documents, then uses its HHEM evaluator to count how often the model invents facts in the summaries. The table reports hallucination rate, factual consistency, answer rate, and average summary length for every entry, and the ranking is refreshed as new models appear. Older versions of the leaderboard are kept around, so you can see how a candidate has trended over time. It is a reference rather than a tool you deploy, but it is a handy way to shortlist a few models before running your own tests, and the interactive version on Hugging Face makes scanning the results easy.