khoaguin/medical-corpus

Introduction

This folder holds the necessary scripts to download the necessary corpora and create indices for more efficient document retrieval for the FedRAG workflow. The corpora listed below were derived from the MedRAG Toolkit [1].

The currently supported corpora are:

Corpus	Size	#Doc.	#Snippets	Domain
StatPearls [3]	~2GBs	9.3k	301.2k	Clinics
Textbooks [4]	~209MBs	18	125.8k	Medicine

For index generation we use the FAISS library.

QA Benchmark Datasets

All the QA benchmark datasets are downloaded from the MIRAGE benchmark [1].

TODOs

Support pubmed, wikipedia, mimic-iv-ext-bhc [6], mimic-iv-note [7].

Prerequisites

Install the just command runner and the uv Python package manager, then create a virtual python env with required packages:

uv sync
source .venv/bin/activate

Running

Available Commands

Command	Description
just	List all available commands
just download	Download raw datasets (statpearls, textbooks, mirage_qa)
just chunk	Chunk the statpearls dataset into smaller pieces
just process	Process statpearls corpus into mock/private datasets with FAISS indices
just process textbooks	Process textbooks corpus into mock/private datasets with FAISS indices
just subset	Create subset datasets for fast experimentation (1% mock, 5% private)
just upload	Upload full processed datasets to Hugging Face
just upload-subsets	Upload subset datasets to Hugging Face

Step-by-Step Workflow

1. Download raw datasets

just download

The raw data for the datasets will be saved into raw_data with the following structure:

raw_data/
├── statpearls/
│   ├── statpearls_NBK430685/
│   └── statpearls_NBK430685.tar.gz
├── textbooks/
│   ├── chunk/
│   ├── .gitattributes
│   └── README.md
└── mirage_qa.json

2. Chunk the statpearls dataset

just chunk

This will process the raw statpearls XML files and create chunked JSONL files in raw_data/statpearls/chunk/. For textbooks, the downloaded raw data are already chunked.

3. Process datasets into mock and private splits

just process statpearls  # for statpearls
just process textbooks   # for textbooks

This will:

• Split the chunked data into mock (1%) and private (99%) datasets
• Build FAISS indices for both splits
• Save everything to processed_data/[corpus_name]/

The processed data structure:

processed_data/
├── statpearls/
│   ├── mock/
│   │   ├── chunk/
│   │   ├── faiss.index
│   │   └── all_doc_ids.npy
│   └── private/
│       ├── chunk/
│       ├── faiss.index
│       └── all_doc_ids.npy
└── textbooks/
    ├── mock/
    │   ├── chunk/
    │   ├── faiss.index
    │   └── all_doc_ids.npy
    └── private/
        ├── chunk/
        ├── faiss.index
        └── all_doc_ids.npy

4. Create subsets for fast experimentation (Optional)

just subset statpearls   # for statpearls
just subset textbooks    # for textbooks

This creates smaller datasets for faster download and experimentation:

• Mock subset: 1% of total data
• Private subset: 5% of total data
• Saved to subsets/[corpus_name]/

The subset data structure:

subsets/
├── statpearls/
│   ├── mock/        (1% of data)
│   │   ├── chunk/
│   │   ├── faiss.index
│   │   ├── all_doc_ids.npy
│   │   └── README.md
│   └── private/     (5% of data)
│       ├── chunk/
│       ├── faiss.index
│       ├── all_doc_ids.npy
│       └── README.md
└── textbooks/
    ├── mock/        (1% of data)
    └── private/     (5% of data)

5. Upload to Hugging Face (Optional)

just upload          # Upload full processed datasets
just upload-subsets  # Upload subset datasets only

Notes

Corpora

To download the data we use the MedRAG Toolkit and the associated MedRAG dataset repository in Hugging Face, which holds all the curated corpus.

According to the privacy policy of StatPearls, the StatPearls content is not allowed to be distributed.
The StatPearls data are downloading directly from the NCBI Bookshelf and the statpearls.py script provided by MedRAG toolkit is used to chunk the data.

The download.py script is responsible for downloading the corpus and executing the required data chunking operations.

FAISS Indexer

We use the FAISS index library for index creation and querying due to its use simplicity and efficient similarity search.

To map the snippets of each corpus to an embedding space and train and create the corpus index, by default we use the all-MiniLM-L6-v2 sentence-transformer model, which maps sentences & paragraphs to a 384 dimensional dense vector space.

The index is created by the ../fedrag/retriever.py script. To build the FAISS indexer, the function build_faiss_index() takes as an input the number of chunks (files) upon which the indexer should be built. If the value is set to 0, then all files are considered during index creation, else only the first [:num_chunks] files are considered. The script also provides a helper function for document retrieval, assuming the index has already been generated.

Specifically, the function query_faiss_index() reads the created index for the requested dataset, converts the input query to its embedded representation and performs a similarity search between the query's embedded vector and the corpus' snippet vectorized formats.

The current implementation of document retrieval for the FAISS index is built with IndexIVFFlat and uses the faiss.METRIC_L2 metric, which means that the lower the score of a retrieved document the better, since L2 Distance measures dissimilarity.

If you want to use another embedding model for generating the embeddings, please change the following properties in the retriever.yaml file.

embedding_model: "sentence-transformers/all-MiniLM-L6-v2"
embedding_dimension: 384

References

1. Xiong, Guangzhi, Qiao Jin, Zhiyong Lu, and Aidong Zhang. "Benchmarking retrieval-augmented generation for medicine." In Findings of the Association for Computational Linguistics ACL 2024, pp. 6233-6251. 2024

2. PubMed corpus was created from articles located at: https://pubmed.ncbi.nlm.nih.gov/

3. StatPearls corpus was created by using 9,330 publicly available StatPearls articles through the NCBI Bookshelf: https://www.ncbi.nlm.nih.gov/books/NBK430685/

4. Textbooks corpus was used in the work of: Jin, Di, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. "What disease does this patient have? a large-scale open domain question answering dataset from medical exams." Applied Sciences 11, no. 14 (2021): 6421. The corpus is also available at: https://github.com/jind11/MedQA

5. Wikipedia corpus was used in the work of: Thakur, Nandan, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. "BEIR: A heterogenous benchmark for zero-shot evaluation of information retrieval models." In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2021). The corpus is also available at: https://huggingface.co/datasets/wikipedia

6. MIMIC-IV-Ext-BHC: Labeled Clinical Notes Dataset for Hospital Course Summarization v1.2.0 (Feb. 3, 2025). https://physionet.org/content/labelled-notes-hospital-course/1.2.0/

7. MIMIC-IV-Note: Deidentified free-text clinical notes v2.2 (Jan. 6, 2023). https://www.physionet.org/content/mimic-iv-note/2.2/

References

• https://github.com/OpenMined/syft-flwr/blob/main/notebooks/fedrag/README.md
• https://github.com/adap/flower/tree/main/examples/fedrag