SFT Reasoning Corpus · OpenMed
Medical-Reasoning-SFT-Mega
A supervised fine-tuning dataset of 1.79 million medical chain-of-thought reasoning conversations across seven models.
What Medical-Reasoning-SFT-Mega is
Medical-Reasoning-SFT-Mega is an English-language supervised fine-tuning corpus created by OpenMed to train instruction-following clinical artificial intelligence models. It aggregates medical question-and-answer interactions equipped with multi-step rationale chains, addressing the need for diverse problem-solving demonstrations across healthcare subjects.
The dataset unifies reasoning outputs generated by seven distinct model architectures into a deduplicated collection of 1,789,998 conversation samples. In total, it comprises approximately 3.78 billion tokens, consisting of roughly 2.22 billion content tokens and 1.56 billion reasoning tokens spanning general medicine, diagnosis, and pharmacology.
How it was made
The corpus was compiled from an initial pool of 2,919,699 samples generated by seven models: MiniMax-M2.1, Baichuan-M3-235B, GPT-OSS-120B, Qwen3-Next-80B, GLM_4.5_Air, Nemotron-Nano-30B, and Trinity-Mini. To prevent individual architectures from dominating repeated clinical prompts, the curators used a fair-distribution deduplication method. The 1,237,711 prompts unique to a single model were retained directly, while the 552,287 shared prompts were iteratively assigned to the model with the lowest allocation of shared questions.
Each record is formatted as a conversational message exchange comprising system, user, and assistant roles. Assistant turns provide both standard answer content and an explicit reasoning_content field containing intermediate chain-of-thought steps, which is populated across 1,789,764 of the final samples.
Key facts
- Total samples
- 1,789,998
- Total tokens
- ~3.78 billion
- Reasoning tokens
- ~1.56 billion
- Unique questions
- 1,237,711
- Contributing models
- 7
- Data format
- Parquet, optimized-parquet
- Language
- English
What you can use it for
- Medical reasoning model fine-tuning
- Train instruction-following language models to produce step-by-step clinical rationales prior to delivering answers.
- Knowledge distillation
- Transfer multi-step reasoning capabilities from large foundation models into compact architectures designed for clinical deployment.
- Comparative reasoning research
- Evaluate and contrast how different model families structure their intermediate reasoning when answering clinical questions.
- Healthcare question-answering development
- Develop question-answering systems capable of addressing clinical, pharmacological, and diagnostic inquiries.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from OpenMed/Medical-Reasoning-SFT-Mega at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 579,117 rows of default/train
messages
list- min
- 2
- median
- 3
- mean
- 2.56
- max
- 4
Strengths and limits
Good at
- Supplies over 1.56 billion tokens of dedicated intermediate chain-of-thought reasoning across nearly 1.79 million samples.
- Employs a fair-distribution deduplication strategy to maintain balanced stylistic contributions from all seven source models.
- Structures conversations with a dedicated reasoning field, separating intermediate thinking from user-facing answers.
Watch out for
- All outputs are generated by artificial intelligence models and may contain medical inaccuracies requiring clinical validation.
- The corpus is not designed for direct clinical decision-making without review by medical professionals.
- The reasoning traces reflect algorithmic approximations rather than verified human clinical decision pathways.
Questions
Can I look inside Medical-Reasoning-SFT-Mega without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is Medical-Reasoning-SFT-Mega under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
How are duplicate questions handled across different source models?
Shared questions are deduplicated via fair distribution, iteratively assigning each duplicate to the contributing model that has received the fewest shared assignments.
What is the structure of each conversation sample?
Samples contain a messages array with system, user, and assistant turns. Assistant messages contain both standard response text and a separate reasoning_content field.
Which models contributed reasoning traces to this dataset?
Data is sourced from MiniMax-M2.1, Baichuan-M3-235B, GPT-OSS-120B, Qwen3-Next-80B, GLM_4.5_Air, Nemotron-Nano-30B, and Trinity-Mini.
What medical topics are represented in the questions?
The questions span clinical medicine, diagnostic workflows, pharmacology, and general medical knowledge.
How to cite it
@dataset{medical_reasoning_sft_mega, title={Medical-Reasoning-SFT-Mega}, author={OpenMed}, year={2025}, publisher={Hugging Face}, url={https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-Mega}}
Written from Medical-Reasoning-SFT-Mega'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.