Medical Reasoning Corpus · miriad
miriad-4.4M
A 4.4-million-pair medical question-answering corpus distilled from peer-reviewed literature for LLM workflows.
What miriad-4.4M is
miriad-4.4M is a large-scale corpus designed for medical reasoning, retrieval, and instruction-tuning workflows. Comprising 4.4 million structured medical question-answer pairs, it represents an earlier release of the Medical Instruction and RetrIeval Dataset (MIRIAD) developed by Qinyue Zheng, Salman Abdullah, Sam Rawal, Cyril Zakka, Sophie Ostmeier, Maximilian Purk, Eduardo Reis, Eric J. Topol, Jure Leskovec, and Michael Moor.
The dataset addresses the need for structured, million-scale medical instruction data by distilling knowledge from peer-reviewed biomedical literature. It provides computational researchers and clinicians with material to evaluate and train language models across biomedical tasks.
The corpus supports a range of downstream applications, including retrieval-augmented generation (RAG), medical information retrieval, hallucination detection, and benchmark evaluations of medical instruction tuning.
How it was made
The dataset was created by extracting source texts from the Semantic Scholar Open Research Corpus (S2ORC) of peer-reviewed biomedical publications. These literature documents were fed as inputs into OpenAI language models to distill structured question-answer pairs.
The resulting data are curated into 4.4 million text-based question-answer pairs and distributed in Parquet format, compatible with standard data-processing libraries such as Hugging Face Datasets, Dask, and Polars.
Key facts
- Size
- 4.4 million question-answer pairs
- Modality
- Text
- Data format
- Parquet
- Source material
- Semantic Scholar Open Research Corpus (S2ORC)
- Generation method
- Distillation using OpenAI language models
- Release year
- 2025
- Authors
- Qinyue Zheng, Salman Abdullah, Sam Rawal, Cyril Zakka, Sophie Ostmeier, Maximilian Purk, Eduardo Reis, Eric J. Topol, Jure Leskovec, Michael Moor
What you can use it for
- Medical Instruction Tuning
- Fine-tuning language models on 4.4 million biomedical question-answer pairs to align them with scientific and medical tasks.
- Retrieval-Augmented Generation (RAG)
- Serving as a structured biomedical knowledge corpus or benchmark for medical retrieval-augmented generation pipelines.
- Hallucination Detection
- Evaluating and improving the factual grounding of medical language models against literature-distilled query-response pairs.
- Medical Information Retrieval
- Benchmarking and training domain-specific information retrieval systems on pairs derived from peer-reviewed literature.
Explore the data
No licence declared
The authors have not published a licence, which means no permission to redistribute the data has been granted to anyone. The rows are not served here; read them on the original dataset page.How the columns are distributed
Over 1,923,243 rows of default/train
answer
string_text- min
- 36
- median
- 474
- mean
- 489.52
- max
- 2,328
paper_id
string_text- min
- 3
- median
- 8
- mean
- 7.79
- max
- 9
paper_title
string_text- min
- 5
- median
- 95
- mean
- 98.89
- max
- 300
paper_url
string_text- min
- 44
- median
- 49
- mean
- 48.79
- max
- 50
passage_position
string_label- 037%
- 117%
- 214%
passage_text
string_text- min
- 103
- median
- 4,522
- mean
- 4581.04
- max
- 9,421
qa_id
string_text- min
- 9
- median
- 15
- mean
- 14.64
- max
- 17
question
string_text- min
- 19
- median
- 97
- mean
- 102.83
- max
- 482
specialty
string_label- Public Health & Epidemiology13%
- Basic Medical Sciences12%
- Pharmacy & Pharmacology8%
venue
string_text- min
- 1
- median
- 28
- mean
- 32.05
- max
- 300
0.37% empty
year
float- min
- 1,793
- median
- 2,014
- mean
- 2011.15
- max
- 2,021
0.00% empty
Strengths and limits
Good at
- Contains 4.4 million question-answer pairs derived from peer-reviewed biomedical publications.
- Provides structured data suitable for diverse downstream tasks including RAG, retrieval, and instruction tuning.
- Available in standard Parquet format compatible with processing frameworks such as Datasets, Polars, and Dask.
Watch out for
- Outputs were generated using OpenAI language models and may inherit model-specific characteristics or errors.
- Must not be used for medical diagnosis, clinical decision-making, or care decisions involving real individuals.
- Has not received review, clearance, or certification as a medical device from any regulatory authority.
Questions
Can I look inside miriad-4.4M without downloading it?
Not here: it declares no licence, so no permission to redistribute its rows has been granted. Its description and statistics are still shown.
What licence is miriad-4.4M under?
Its authors have not declared a licence, which means no permission to reuse or redistribute it has been granted.
What primary source texts were used to build miriad-4.4M?
The dataset was built using biomedical literature documents from the Semantic Scholar Open Research Corpus (S2ORC).
Can this dataset be used in clinical practice?
No. The dataset is provided strictly for academic research and educational exploration, and must not be used for medical diagnosis, clinical decisions, or as a certified medical device.
How are the questions and answers generated?
S2ORC biomedical documents were provided as inputs to OpenAI language models to distill structured question-answer pairs.
How to cite it
Qinyue Zheng, Salman Abdullah, Sam Rawal, Cyril Zakka, Sophie Ostmeier, Maximilian Purk, Eduardo Reis, Eric J. Topol, Jure Leskovec, Michael Moor. MIRIAD: Augmenting LLMs with millions of medical query-response pairs. arXiv:2506.06091 [cs.CL], 2025.
Written from miriad-4.4M's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.