Skip to content

Dataset · google

FRAMES benchmark for RAG systems

google/frames-benchmark

824 complex multi-hop questions across 2–15 Wikipedia articles to test RAG factuality, retrieval, and reasoning.

What's inside

FRAMES is an evaluation dataset of 824 questions requiring information from 2–15 Wikipedia articles. It evaluates RAG systems on factuality, retrieval precision, and reasoning, covering diverse topics. The language is English, with a volume of under 1,000 examples; details are available in the arXiv paper.

What you can do with it

  • Evaluating RAG pipelines
  • Testing LLMs on multi-hop QA
  • Comparing retrieval components
Tasks
Text classification, Token classification, Table question answering, Question answering
Languages
English
Size
up to 1K rows
Downloads this month
31.1K
Likes
276

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("google/frames-benchmark")