README.md

metadata

annotations_creators:
  - machine-generated
language:
  - en
license: cc-by-4.0
tags:
  - medical
  - rag
  - synthetic-qa
  - lay-symptom
pretty_name: MediMaven-QA v1.0
size_categories:
  - 100K<n<1M
dataset_info:
  - config_name: kb_chunks
    features:
      - name: id
        dtype: string
      - name: url
        dtype: string
      - name: title
        dtype: string
      - name: section
        dtype: string
      - name: source
        dtype: string
      - name: text
        dtype: string
      - name: retrieved_date
        dtype: string
      - name: n_tokens
        dtype: int64
    splits:
      - name: train
        num_bytes: 133140842
        num_examples: 70743
    download_size: 51361461
    dataset_size: 133140842
  - config_name: qa_long
    features:
      - name: chunk_id
        dtype: string
      - name: question
        dtype: string
      - name: answer
        dtype: string
    splits:
      - name: train
        num_bytes: 52621793
        num_examples: 143280
    download_size: 26138154
    dataset_size: 52621793
  - config_name: qa_wide
    features:
      - name: chunk_id
        dtype: string
      - name: qa
        list:
          - name: answer
            dtype: string
          - name: question
            dtype: string
    splits:
      - name: train
        num_bytes: 49971385
        num_examples: 70018
    download_size: 27339393
    dataset_size: 49971385
configs:
  - config_name: kb_chunks
    data_files:
      - split: train
        path: kb_chunks/train-*
  - config_name: qa_long
    data_files:
      - split: train
        path: qa_long/train-*
  - config_name: qa_wide
    data_files:
      - split: train
        path: qa_wide/train-*

🩺 MediMaven-QA v1.1

MediMaven-QA is a multi-source, medically-grounded question-answer (QA) corpus designed for retrieval-augmented generation (RAG), LLM fine-tuning, and benchmarking factual consistency in healthcare chatbots. Version 1.1 aggregates Mayo Clinic, NHS.uk, WebMD, and CDC consumer health pages, then auto-generates lay-friendly Q&A pairs with GPT-4o. Each passage is PHI-scrubbed, CC-BY licensed, and versioned with Weights & Biases.

• 150 k QA pairs

• 28 M tokens

• Balanced mix of concise FAQs and narrative symptom descriptions

📦 Dataset Contents

Config (name)	Rows	What it holds	Typical use-case
chunks	70 243	400-token, sentence-aware context windows with rich metadata (id, url, title, section, source, n_token, text)	RAG context store / retriever training
qa_wide	70 018	List-of-dict QA per chunk_id → single row may have ≥1 QA pair	Fast retrieval + generation, keeps chunk linkage
qa_long	143 221	Fully exploded (chunk_id, question, answer)	Classic supervised QA fine-tuning or eval

⚠️ Disclaimer — This corpus is for research & benchmarking only.
It is not a diagnostic tool and should not be used in clinical workflows.

🚀 Quick Load

from datasets import load_dataset

# pick one of these configs
qa_long  = load_dataset("bernard-kyei/medimaven-qa-data", "qa_long", split="train")
qa_wide  = load_dataset("bernard-kyei/medimaven-qa-data", "qa_wide", split="train")
# accompany with chunks to get contexts
chunks   = load_dataset("bernard-kyei/medimaven-qa-data", "kb_chunks",  split="train")

print(qa_long[0]["question"])
print(qa_long[0]["answer"])

🛠️ Creation Pipeline

1. Scraped Mayo Clinic, NHS.uk, WebMD, (licensed for public use).
2. Sentence-aware chunking (spaCy).
3. GPT-4o-mini synthetic Q-A:
   • one concise lay question
   • one symptom-narrative question (e.g., “I woke up with a painful lump …”)
4. Cost: $40 for ~150 k pairs.
5. Versions tracked as W&B artifacts.

Stage	Tooling	Notes
1️⃣ Crawl	Scrapy + Splash	Mayo Clinic, NHS.uk, WebMD, (public-domain / permissive T&Cs)
2️⃣ Chunk	spaCy sentenciser	≈400 tokens / chunk; keeps heading context
3️⃣ Synthetic QA	GPT-4o-mini (gpt-4o-mini-2024-05-preview)	• 1 concise lay Q • 1 symptom-narrative Q → cost $40 for ~150 k pairs
4️⃣ Versioning	Weights & Biases Artifacts	kb_chunks, qa_wide qa_long

📊 Key Stats

Metric	Value
Total context tokens	~28 M
Avg. tokens / chunk	390
Unique host domains	4
QA pairs / chunk (mean)	2.0
% symptom-narrative Qs	51 %

🧩 Dataset Structure

Field	Type	Description
id	string	UUID4 per chunk
url	string	Canonical page URL
title	string	Page headline
section	string	<h2>/<h3> section header
source	string	Domain slug (mayo, nhs, webmd, cdc)
text	string	≤ 400-token chunk
n_tokens	int32	Token count (tiktoken/gpt-4 tokenizer)
qa	list[dict]	Each dict has question, answer

🗒️ Changelog

Date (UTC)	Version	Notes
2025-05-29	1.1	Multi-source crawl, synthetic QA, PHI scrub, W&B tracking
2024-11-01	1.0	MedQuad + iCliniq seed release

Contributions

PRs welcome! Please open an issue describing planned changes.
We follow the Hugging Face Datasets community guidelines.

Maintainer: Bernard Kyei-Mensah — [email protected]
LinkedIn / GitHub: @dranreb1660

Acknowledgements

Thanks to the open-source healthcare community, Hugging Face, and contributors who reported issues and suggested improvements.

📜 Citation

@misc{KyeiMensah2025MediMavenQA,
  author  = {Kyei-Mensah, Bernard},
  title   = {MediMaven-QA: A Citation-Preserving Medical Q\A Dataset with Symptom Narratives},
  year    = {2025},
  url     = {https://huggingface.co/datasets/dranreb1660/medimaven-qa-data},
  note    = {Version 1.0}
}