metadata

license: apache-2.0

FRAMES: Factuality, Retrieval, And reasoning MEasurement Set

FRAMES is a comprehensive evaluation dataset designed to test the capabilities of Retrieval-Augmented Generation (RAG) systems across factuality, retrieval accuracy, and reasoning.

Dataset Overview

824 challenging multi-hop questions requiring information from 2-15 Wikipedia articles
Questions span diverse topics including history, sports, science, animals, health, etc.
Each question is labeled with reasoning types: numerical, tabular, multiple constraints, temporal, and post-processing
Gold answers and relevant Wikipedia articles provided for each question

Key Features

Tests end-to-end RAG capabilities in a unified framework
Requires integration of information from multiple sources
Incorporates complex reasoning and temporal disambiguation
Designed to be challenging for state-of-the-art language models

Usage

This dataset can be used to:

Evaluate RAG system performance
Benchmark language model factuality and reasoning
Develop and test multi-hop retrieval strategies

Baseline Results

We provide baseline results using state-of-the-art models like Gemini-Pro-1.5-0514:

Naive prompting: 40.8% accuracy
BM25 retrieval (4 docs): 47.4% accuracy
Oracle retrieval: 72.9% accuracy
Multi-step retrieval & reasoning: 66% accuracy

Citation

If you use this dataset in your research, please cite our paper:

We hope FRAMES will be useful for advancing RAG systems and language model capabilities. For more details, please refer to our full paper.