---
dataset_info:
  features:
  - name: source
    dtype: string
  - name: query_id
    dtype: string
  - name: query
    dtype: string
  - name: doc_name
    dtype: string
  - name: answer
    sequence: string
  - name: doc_url
    dtype: string
  - name: num_doc_labels
    dtype: int32
  - name: doc_pool
    sequence:
    - name: mapped_id
      dtype: string
    - name: doc_name
      dtype: string
    - name: doc_chunk
      dtype: string
    - name: support
      dtype: int32
  - name: oracle
    struct:
    - name: mapped_id
      dtype: string
    - name: doc_name
      dtype: string
    - name: doc_chunk
      dtype: string
    - name: support
      dtype: int32
  splits:
  - name: train
    num_bytes: 59957279
    num_examples: 7560
  download_size: 35533441
  dataset_size: 59957279
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: apache-2.0
task_categories:
- text-generation
language:
- en
pretty_name: mirage
size_categories:
- 1K<n<10K
---

# MIRAGE dataset

MIRAGE is a benchmark dataset for evaluating Retrieval-Augmented Generation (RAG) systems, featuring 7,560 QA pairs and 37,800 context pools curated from diverse Wikipedia-based QA datasets (IfQA, NaturalQA, TriviaQA, DROP, PopQA). MIRAGE enables robust assessment of LLMs and retrievers under realistic, noisy, and oracle settings, and introduces novel metrics for analyzing context sensitivity, noise vulnerability, and retrieval effectiveness.

You can find our paper and github repository below:
[![Read the Paper](https://img.shields.io/badge/Paper-arXiv%3A2054.17137-B31B1B.svg)](https://arxiv.org/abs/2504.17137) [![GitHub Repo](https://img.shields.io/badge/GitHub-MIRAGE-blue?logo=github)](https://github.com/nlpai-lab/MIRAGE)