--- license: apache-2.0 task_categories: - question-answering - text-generation - text-retrieval language: - en tags: - medical - healthcare - clinical - medical-qa - medical-llm - healthcare-ai - doctor - patient - diagnosis - treatment - drugs - pharmacology - instruction-tuning - fine-tuning - sft - rag - chatdoctor - medgenius - llama - mistral - gpt - "2026" - biomedical pretty_name: Medical Intelligence 76k - 2026 Edition size_categories: - 10K **Production-ready medical AI dataset for training diagnosis, treatment reasoning, and doctor-patient conversational systems.** > > 76,000 **engineered** (not collected) English Q&A pairs โ€” covering 620+ diseases, > 438+ FDA-approved drugs, and real patient-doctor conversations. Built with a > 5-stage quality pipeline. Commercial-safe **Apache 2.0**. [![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](https://opensource.org/licenses/Apache-2.0) ![Rows](https://img.shields.io/badge/rows-76%2C000-success) ![Quality](https://img.shields.io/badge/zero--issue--audit-passed-brightgreen) ![Language](https://img.shields.io/badge/language-100%25_English-informational) ![Updated](https://img.shields.io/badge/updated-2026-orange) ![Source-traceable](https://img.shields.io/badge/provenance-row--level-blueviolet) **Created by [Huzefa Nalkheda Wala](https://huzefanalkhedawala.in/)** โ€” AI Product Engineer & Medical AI Researcher ยท Creator of the original [Medical Intelligence Dataset 40k](https://huggingface.co/datasets/huzaifa525/Medical_Intelligence_Dataset_40k_Rows_of_Disease_Info_Treatments_and_Medical_QA) (DOI: [10.57967/hf/3751](https://doi.org/10.57967/hf/3751)) and the [MedGenius LLaMA-3.2B](https://huggingface.co/huzaifa525/MedGenius_LLaMA-3.2B) medical LLM family. --- ## โœจ This dataset was engineered, not collected. Most "medical Q&A" datasets on HuggingFace are scraped dumps with HTML, formula artifacts, language drift, and 5-15% duplicate rows. They get you to a demo quickly โ€” and break in production. **This one is different. Every row went through a 5-stage quality pipeline:** | Stage | What it does | Rows dropped | |-------|--------------|--------------| | 1. MinHash LSH dedup | Remove fuzzy duplicates against v1 (Jaccard 0.85) | 3,290 | | 2. CSV-artifact filter | Strip `#NAME?`, `#REF!`, formula injections | 38 | | 3. HTML / entity cleanup | Remove `

`, `&`, escape sequences | 55+ | | 4. Language filter | Drop Spanish hallucinations + non-English topics | 7,769 | | 5. Length & exact-dup filter | Min 10 chars, no duplicate inputs | 1,463 | **Final post-clean audit (run on every row):** - โœ… 0 null/empty rows - โœ… 0 Spanish content - โœ… 0 CSV artifacts - โœ… 0 HTML tags - โœ… 0 code fences - โœ… 0 duplicate inputs - โœ… 100% English (verified) **This is the difference between a dataset that demos and one that ships.** --- ## ๐Ÿ†• What changed from v1 โ†’ v2? | Aspect | v1 (40k, 2024) | v2 (76k, 2026) | |--------|----------------|----------------| | **Rows** | 40,442 | **76,000** (+90%) | | **Topics covered** | ~42 specialties | **620+ conditions, 438+ drugs** | | **Drug coverage** | Limited | **1,500 latest FDA labels (2025-2026)** | | **Currency** | Pre-2024 | **2026-refreshed** (current FDA approvals) | | **Provenance tagging** | Single source | **Row-level `source` column** (every answer traceable) | | **Quality pipeline** | Basic | **5-stage engineered** (dedup, language, artifacts, length, duplicates) | | **Spanish content** | Some present | **0 (verified)** | | **CSV artifacts** | 52 `#NAME?` rows | **0** | | **HTML leakage** | Some `

` tags | **0** | | **License** | Apache 2.0 | Apache 2.0 (preserved) | If you used v1, **v2 is a strict upgrade** โ€” same schema, 2x rows, cleaner data, current drug knowledge. --- ## ๐Ÿ“Š What's inside | Category | Coverage | |----------|----------| | ๐Ÿฆ  **Diseases & conditions** | 620+ topics: Anemia, Malaria, Meningitis, Edema, Bursitis, Leishmaniasis, Chikungunya, and many more | | ๐Ÿ’Š **Drugs & medications** | 438+ FDA-approved drugs with patient-style Q&A on indications, dosing, side effects, contraindications, interactions | | ๐Ÿ‘จโ€โš•๏ธ **Patient-doctor dialogues** | Real-world consultation patterns and clinical reasoning | | ๐Ÿ“š **Exam-style explanations** | AIIMS/NEET-PG MCQ derivations with detailed expert explanations | --- ## ๐Ÿ“‹ Example rows (one from each source) **Patient-doctor dialogue (`v1`)** ``` INPUT: How many people are affected by neuropathy, ataxia, and retinitis pigmentosa? OUTPUT: The prevalence of NARP is unknown. This disorder is probably less common than a similar but more severe condition, Leigh syndrome, which affects about 1 in 40,000 people. ``` **Exam-style reasoning (`medmcqa`)** ``` INPUT: The number of deaths prevented as a result of a particular health programme is best evaluated by: Options: A. Cost accounting B. Cost benefit analysis C. Cost effective analysis D. None of the above OUTPUT: The correct answer is C. Cost effective analysis. Explanation: Cost effective analysis evaluates the benefits of a particular health programme in terms of results achieved, like number of lives saved or number of days free from disease. Ref: Park 22nd edition pg: 814. ``` **Drug information (`openFDA`)** ``` INPUT: What is the recommended starting dose of mirtazapine and how should it be taken? OUTPUT: The recommended starting dose for mirtazapine tablets, USP is 15 mg per day, administered in a single dose, preferably in the evening prior to sleep. ``` **Patient health education (`MedlinePlus`)** ``` INPUT: What types of preventive care are included in a regular checkup? OUTPUT: Preventive care may include screening tests, vaccines, and education and counseling about lifestyle choices such as diet, exercise, and use of alcohol, tobacco, or other substances. ``` Notice the consistency โ€” clean prose, no HTML, no formula artifacts, grounded in real source material. **Every row meets this bar.** **Question styles included:** - "What is X?" / "What causes X?" / "What are the symptoms of X?" - "How is X diagnosed/treated/prevented?" - "Doctor, I haveโ€ฆ" โ€” patient self-report consultations - "Can I take X with Y?" โ€” drug interaction queries - "What should I do ifโ€ฆ?" โ€” clinical decision support --- ## ๐Ÿš€ Quickstart ### Load with ๐Ÿค— Datasets ```python from datasets import load_dataset ds = load_dataset("huzaifa525/Medical_Intelligence_Dataset_76k_2026_Edition") print(ds["train"][0]) # {"input": "...", "output": "...", "source": "openFDA:Ibuprofen"} ``` ### Fine-tune (axolotl / TRL example) ```python from datasets import load_dataset ds = load_dataset("huzaifa525/Medical_Intelligence_Dataset_76k_2026_Edition", split="train") def to_chat(row): return { "messages": [ {"role": "user", "content": row["input"]}, {"role": "assistant", "content": row["output"]}, ] } ds = ds.map(to_chat, remove_columns=ds.column_names) # Plug into TRL's SFTTrainer, axolotl, llama-factory, Unsloth, etc. ``` > ๐Ÿ“– **Want a step-by-step walkthrough?** Read the author's blog: > [Fine-Tuning LLaMA 3.2 for Medical AI: Building MedGenius from 40K+ Dialogues](https://huzefanalkhedawala.in/blog/2025/01/10/fine-tuning-llama-medical-ai/) > โ€” covers LoRA configs, Unsloth setup, evaluation, and how the original 40k was used to fine-tune a model that hit 89% medical accuracy for $473 in 8.4 hours on a single RTX 4090. ### Use for RAG The `source` column traces every answer back to its origin (FDA label, MedlinePlus topic, etc.) โ€” perfect for building auditable retrieval pipelines. > ๐Ÿ“– **Production RAG playbook:** > [How to Build Production RAG Systems](https://huzefanalkhedawala.in/blog/2025/01/15/how-to-build-production-rag-systems/) > โ€” Qdrant + BGE + hybrid search recipe used to serve 50K+ daily queries at 850ms p95. --- ## ๐Ÿ“ฆ Schema | Column | Type | Description | |--------|------|-------------| | `input` | string | Question, patient query, or clinical scenario | | `output` | string | Grounded answer with explanation | | `source` | string | Provenance tag: `v1`, `medmcqa`, `openFDA:`, `MedlinePlus:` | --- ## ๐Ÿ—๏ธ Source distribution | Source | Rows | License | Description | |--------|------|---------|-------------| | **v1** (original) | 38,893 | Apache 2.0 | Patient-doctor Q&A and disease information from the original 40k release | | **medmcqa** | 19,640 | Apache 2.0 | Medical MCQs (AIIMS / NEET-PG) converted to Q&A format with expert explanations | | **openFDA** | 10,156 | Public domain | Grounded Q&A synthesized from **1,500 latest FDA drug labels (2025-2026)** | | **MedlinePlus** | 7,311 | Public domain (NIH) | Grounded Q&A synthesized from English health topic articles | | **Total** | **76,000** | Apache 2.0 | Fully commercial-safe | --- ## ๐Ÿ”ง Synthesis details New rows were generated with `openai/gpt-oss-120b` via OpenRouter using **strict grounding** โ€” every answer must be supported by the provided source text, no free-form invention allowed. This drastically cuts hallucination versus pure prompt-only synthesis. The full quality pipeline is described at the top of this card. --- ## ๐Ÿ“Š Stats at a glance | Metric | Value | |--------|-------| | **Rows** | 76,000 | | **Total text** | 52.8 MB | | **Unique source tags** | 1,060 | | **Median input length** | 105 chars | | **Median output length** | 313 chars | | **p95 input** | 546 chars | | **p95 output** | 1,745 chars | | **Language coverage** | 100% English | | **Splits** | single `train` split | --- ## ๐ŸŽฏ Built for production. Use it for: - ๐Ÿค– **Fine-tuning medical LLMs** for Llama / Mistral / Qwen / Gemma / Phi โ€” SFT-ready out of the box - ๐Ÿ’ฌ **Healthcare chatbots & triage assistants** โ€” patient-style phrasing, doctor-style answers - ๐Ÿ” **Medical RAG systems** โ€” row-level provenance tagging makes every answer auditable - ๐Ÿฉบ **Clinical decision support** โ€” grounded Q&A backed by FDA labels, MedlinePlus, and exam-grade reasoning - ๐Ÿ’Š **Drug Q&A & pharmacovigilance bots** โ€” current 2025-2026 FDA approvals, indications, contraindications - ๐Ÿงช **Medical instruction-tuning research** โ€” large enough for production SFT, small enough for fast iteration - ๐ŸŽ“ **Medical education tools** โ€” exam prep with expert explanations This is not a learning-toy dataset. It's the foundation you ship a medical AI product on. --- ## ๐Ÿ“ˆ Benchmark โ€” does fine-tuning on this data actually work? **Yes โ€” and here's the receipt.** The original v1 (40k subset of this same dataset) was used to fine-tune [MedGenius LLaMA-3.2B](https://huggingface.co/huzaifa525/MedGenius_LLaMA-3.2B) (LoRA r=16, alpha=32, Unsloth, 8.4 hrs on a single RTX 4090, total cost $473). | Metric | Base LLaMA-3.2 | MedGenius (v1-trained) | Improvement | |--------|----------------|------------------------|-------------| | **ROUGE-1** | 0.42 | **0.78** | +86% | | **ROUGE-L** | 0.38 | **0.72** | +89% | | **Medical Accuracy** | 61% | **89%** | +46% | | **Empathy Score** (5-pt) | 3.2 | **4.7** | +47% | **v2 (this dataset) is a strict superset** โ€” same v1 data, plus 37k cleaned and deduped rows from openFDA + MedlinePlus + medmcqa. Models trained on v2 are expected to outperform v1-trained models, especially on: - **Drug knowledge questions** (10k+ openFDA rows on 2025-2026 approvals) - **Patient-friendly explanations** (7k+ MedlinePlus rows) - **Exam-style clinical reasoning** (19k+ medmcqa rows with expert explanations) > **v2 fine-tune benchmarks coming soon.** Want to be the first to publish > them? Train your model on this dataset and tag the author โ€” top results > will be featured here. --- ## ๐Ÿค Companion models by the same author Already-published models trained on the v1 (40k) version of this dataset โ€” pair them with this 76k edition for an even stronger baseline: | Model | Params | Quantized | Use case | |-------|--------|-----------|----------| | [MedGenius_LLaMA-3.2B](https://huggingface.co/huzaifa525/MedGenius_LLaMA-3.2B) | 3B | F16 | Lightweight medical chatbot, edge inference | | [MedGenius_LLaMA-3.2B-Q4_K_M-GGUF](https://huggingface.co/huzaifa525/MedGenius_LLaMA-3.2B-Q4_K_M-GGUF) | 3B | 4-bit GGUF | `llama.cpp` / Ollama / LM Studio deployment | | [Doctoraifinetune-3.1-8B](https://huggingface.co/huzaifa525/Doctoraifinetune-3.1-8B) | 8B | 4-bit | Disease diagnosis, symptom analysis, treatment recommendations | | [Doctoraifinetune-3.1-8B-gguf](https://huggingface.co/huzaifa525/Doctoraifinetune-3.1-8B-gguf) | 8B | GGUF | Local deployment of the 8B model | **MedGenius LLaMA-3.2B benchmarks** (vs base LLaMA 3.2): - ROUGE-1: **0.78** (vs 0.42) ยท ROUGE-L: **0.72** (vs 0.38) - Medical Accuracy: **89%** (vs 61%) ยท Empathy Score: **4.7/5** (vs 3.2/5) --- ## ๐Ÿ“œ Citation If this dataset helps your work, please cite both the dataset and the upstream sources: ```bibtex @dataset{nalkhedawala2026medical76k, author = {Huzefa Nalkheda Wala}, title = {Medical Intelligence Dataset - 76k - 2026 Edition}, year = {2026}, publisher = {Hugging Face}, url = {https://huggingface.co/datasets/huzaifa525/Medical_Intelligence_Dataset_76k_2026_Edition}, note = {Apache 2.0. Built on the original 40k Medical Intelligence Dataset (DOI: 10.57967/hf/3751) and expanded with grounded synthesis from openFDA and MedlinePlus.} } @dataset{nalkhedawala2024medical40k, author = {Huzefa Nalkheda Wala}, title = {Medical Intelligence Dataset - 40k Rows of Disease Info, Treatments and Medical QA}, year = {2024}, publisher = {Hugging Face}, doi = {10.57967/hf/3751}, url = {https://huggingface.co/datasets/huzaifa525/Medical_Intelligence_Dataset_40k_Rows_of_Disease_Info_Treatments_and_Medical_QA} } @inproceedings{pal2022medmcqa, title = {MedMCQA: A Large-scale Multi-Subject Multi-Choice Dataset for Medical domain Question Answering}, author = {Pal, Ankit and Umapathi, Logesh Kumar and Sankarasubbu, Malaikannan}, booktitle = {Conference on Health, Inference, and Learning}, pages = {248--260}, year = {2022}, organization = {PMLR} } ``` - **openFDA** โ€” https://open.fda.gov/ - **MedlinePlus** โ€” https://medlineplus.gov/ --- ## โš–๏ธ License **Apache 2.0** โ€” every contributing source is either Apache-licensed or U.S. public domain. Use it commercially, modify it, redistribute it. Attribution to the upstream sources is appreciated. --- ## โš ๏ธ Medical disclaimer This dataset is provided for **research and educational purposes only**. Content must not be used as a substitute for professional medical advice, diagnosis, or treatment. Always consult a qualified healthcare provider for any medical concerns. LLM-synthesized rows are grounded in source material but may still contain errors โ€” always verify critical information against authoritative sources. --- ## ๐Ÿ‘ค About the author **Huzefa Nalkheda Wala** is an AI Product Engineer and Medical AI Researcher based in India, currently building production AI systems at [Ruby CRM / CleverFlow](https://www.rubycrm.ai/) (Dubai). He previously won the **IBM World Hack Challenge (2020)** and **1st place at Vadodara Startup Festival (2022)** for a patented medical IoT device (Indian Design Patent No. 375474-001). His open-source medical AI work โ€” including the original 40k Medical Intelligence Dataset, the MedGenius LLaMA-3.2B family, and DoctorAI 8B โ€” has crossed **1,100+ combined downloads** across Hugging Face and Kaggle. ### Connect - ๐ŸŒ **Website:** [huzefanalkhedawala.in](https://huzefanalkhedawala.in/) - ๐Ÿ’ผ **LinkedIn:** [linkedin.com/in/huzefanalkheda](https://www.linkedin.com/in/huzefanalkheda/) (3,000+ followers) - ๐Ÿ™ **GitHub:** [github.com/huzaifa525](https://github.com/huzaifa525) - ๐Ÿค— **HuggingFace:** [huggingface.co/huzaifa525](https://huggingface.co/huzaifa525) - ๐Ÿ“Š **Kaggle:** [kaggle.com/huzefanalkheda](https://www.kaggle.com/huzefanalkheda) - ๐Ÿ“ **Medium:** [huzefanalkheda.medium.com](https://huzefanalkheda.medium.com/) - ๐ŸŽ“ **ORCID:** [0009-0000-1455-6745](https://orcid.org/0009-0000-1455-6745) - ๐• **Twitter:** [@HuzaifaNalkheda](https://x.com/HuzaifaNalkheda) - โœ‰๏ธ **Email:** huzaifanalkhedaemp@gmail.com > *"Two hours of scrolling won't change your life. Two hours of building might."* > โ€” Huzefa Nalkheda Wala --- ## ๐Ÿ™ Acknowledgements Built on the shoulders of the medical AI community: - The **MedMCQA team** (Pal, Umapathi, Sankarasubbu) for releasing high-quality MCQ data under Apache 2.0 - **openFDA** and the U.S. FDA for keeping drug-label data publicly accessible - The **NIH MedlinePlus team** for an exemplary public-domain health knowledge base - Everyone who used, starred, or cited the original 40k v1 โ€” your feedback shaped this v2 **If you build something cool with this, drop a ๐Ÿ‘ like, leave a comment, and tag me on [LinkedIn](https://www.linkedin.com/in/huzefanalkheda/) โ€” I'd love to feature your work.**