Welcome to the GitHub page of MIRAGE
(Medical Information Retrieval-Augmented Generation Evaluation) Benchmark! This repository contains a comprehensive dataset and benchmark results aimed at evaluating Retrieval-Augmented Generation (RAG) systems for medical question answering (QA). We use the MedRAG
toolkit to evaluate existing solutions of various components in RAG on MIRAGE.
The benchmark data is stored as benchmark.json
in this repo, which can also be downloaded from Google Drive.
Snippet ids of the top 10k snippets for each task in MIRAGE
retrieved by all retrievers in MedRAG
can be downloaded with
wget -O retrieved_snippets_10k.zip https://virginia.box.com/shared/static/cxq17th6eisl2pn04vp0x723zczlvlzc.zip
For a realistic evaluation purpose, MIRAGE adopts four key evaluation settings:
Zero-Shot Learning (ZSL): Input QA systems are evaluated in a zero-shot setting where in-context few-shot learning is not permitted.
Multi-Choice Evaluation (MCE): Multi-choice Questions are used to evaluate given systems.
Retrieval-Augmented Generation (RAG): Input systems should perform retrieval-augmented generation, which need to collect external information for accurate and reliable answer generation.
Question-Only Retrieval (QOR): To align with real-world cases of medical QA, answer options should not be provided as input during the retrieval.
The following figure presents the overview of MIRAGE, which shows that MIRAGE contains five commonly used datasets for medical QA for the evaluation of RAG systems, including three medical examination QA datasets and two biomedical research QA datasets:
- MMLU-Med: A medical examination QA dataset with 1089 questions. A subset of six tasks that are related to biomedicine are selected from MMLU, including anatomy, clinical knowledge, professional medicine, human genetics, college medicine, and 996 college biology.
- MedQA-US: A medical examination QA dataset. We focus on the real-world English subset in MedQA with questions from the US Medical Licensing Examination (MedQA-US), including 1273 four-option test samples.
- MedMCQA: A medical examination QA dataset. We chose the dev set of the original MedMCQA, which includes 4183 medical questions from Indian medical entrance exams.
- PubMedQA*: A biomedical research QA dataset. We build PubMedQA* by removing given contexts in the 500 expert-annotated test samples of PubMedQA. The possible answer to a PubMedQA* question can be yes/no/maybe, reflecting the authenticity of the question statement based on scientific literature.
- BioASQ-Y/N: A biomedical research QA dataset. We select the Yes/No questions in the ground truth test set of BioASQ Task B from the most recent five years (2019-2023), including 618 questions in total. The ground truth snippets are removed in this benchmark.
Statistics of datasets in MIRAGE are shown below:
Dataset | Size | #O. | Avg. L | Source |
---|---|---|---|---|
MMLU-Med | 1,089 | 4 | 63 | Examination |
MedQA-US | 1,273 | 4 | 177 | Examination |
MedMCQA | 4,183 | 4 | 26 | Examination |
PubMedQA* | 500 | 3 | 24 | Literature |
BioASQ-Y/N | 618 | 2 | 17 | Literature |
(#O.: numbers of options; Avg. L: average token counts in each question.)
The following table shows the benchmark results of different backbone LLMs.
This table shows the comparison of different corpora and retrievers on MIRAGE.
Load the benchmark:
>>> import json
>>> benchmark = json.load(open("benchmark.json"))
Load specific datasets in the benchmark (e.g., mmlu):
>>> from src.utils import QADataset
>>> dataset_name = "mmlu"
>>> dataset = QADataset(dataset_name)
>>> print(len(dataset))
1089
>>> print(dataset[0])
{'question': 'A lesion causing compression of the facial nerve at the stylomastoid foramen will cause ipsilateral', 'options': {'A': 'paralysis of the facial muscles.', 'B': 'paralysis of the facial muscles and loss of taste.', 'C': 'paralysis of the facial muscles, loss of taste and lacrimation.', 'D': 'paralysis of the facial muscles, loss of taste, lacrimation and decreased salivation.'}, 'answer': 'A'}
Evaluate prediction results saved in ./prediction
for both CoT generation and RAG with 32 snippets:
# CoT with GPT-3.5
python src/evaluate.py --results_dir ./prediction --llm_name OpenAI/gpt-35-turbo-16k
# MedRAG-32 with GPT-3.5
python src/evaluate.py --results_dir ./prediction --llm_name OpenAI/gpt-35-turbo-16k --rag --k 32
# CoT with GPT-4
python src/evaluate.py --results_dir ./prediction --llm_name OpenAI/gpt-4-32k
# MedRAG-32 with GPT-4
python src/evaluate.py --results_dir ./prediction --llm_name OpenAI/gpt-4-32k --rag --k 32
To submit results of your new system on the Leaderboard, please send an email to Guangzhi Xiong ([email protected]) with
- The name of your system and its components
- Performance of the system on different subtasks & Average performance
- A reference link to your results
@inproceedings{xiong-etal-2024-benchmarking,
title = "Benchmarking Retrieval-Augmented Generation for Medicine",
author = "Xiong, Guangzhi and
Jin, Qiao and
Lu, Zhiyong and
Zhang, Aidong",
editor = "Ku, Lun-Wei and
Martins, Andre and
Srikumar, Vivek",
booktitle = "Findings of the Association for Computational Linguistics ACL 2024",
month = aug,
year = "2024",
address = "Bangkok, Thailand and virtual meeting",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.findings-acl.372",
doi = "10.18653/v1/2024.findings-acl.372",
pages = "6233--6251",
abstract = "While large language models (LLMs) have achieved state-of-the-art performance on a wide range of medical question answering (QA) tasks, they still face challenges with hallucinations and outdated knowledge. Retrieval-augmented generation (RAG) is a promising solution and has been widely adopted. However, a RAG system can involve multiple flexible components, and there is a lack of best practices regarding the optimal RAG setting for various medical purposes. To systematically evaluate such systems, we propose the Medical Information Retrieval-Augmented Generation Evaluation (MIRAGE), a first-of-its-kind benchmark including 7,663 questions from five medical QA datasets. Using MIRAGE, we conducted large-scale experiments with over 1.8 trillion prompt tokens on 41 combinations of different corpora, retrievers, and backbone LLMs through the MedRAG toolkit introduced in this work. Overall, MedRAG improves the accuracy of six different LLMs by up to 18{\%} over chain-of-thought prompting, elevating the performance of GPT-3.5 and Mixtral to GPT-4-level. Our results show that the combination of various medical corpora and retrievers achieves the best performance. In addition, we discovered a log-linear scaling property and the {``}lost-in-the-middle{''} effects in medical RAG. We believe our comprehensive evaluations can serve as practical guidelines for implementing RAG systems for medicine.",
}