Jianwen/MedVerse14k

--- license: apache-2.0 task_categories: - question-answering - text-generation language: - en tags: - medical - reasoning - chain-of-thought - dag - petri-net - medqa pretty_name: MedVerse14k size_categories: - 10K<n<100K --- ## MedVerse14k ### Overview MedVerse14k is the training dataset for [MedVerse: Efficient and Reliable Medical Reasoning via DAG-Structured Parallel Execution](https://arxiv.org/abs/2602.07529). It contains **13,904 medical multiple-choice questions** annotated with knowledge-grounded DAG reasoning paths generated by the MedVerse Curator pipeline. ### Dataset Summary | Split | Examples | Description | |-------|----------|-------------| | `train` | 12,513 | Medical QA with DAG reasoning paths, Petri-net plans, and step-by-step CoT execution | | `validation` | 1,391 | Held-out split (10%) | Source questions are drawn from [UCSC-VLAA/MedReason](https://huggingface.co/datasets/UCSC-VLAA/MedReason), which aggregates MedMCQA, MedQA, PubMedQA, MedXpertQA, and other medical QA benchmarks. ### Data Format Each example contains the following fields: | Field | Description | |-------|-------------| | `id` | Unique sample identifier | | `Question` | Medical multiple-choice question stem | | `Options` | Answer choices (A/B/C/D) | | `Original Reason Path` | Knowledge-grounded A→B→C reasoning chains | | `Transient Path` | DAG path structure among transient steps | | `Transient Plan Prompt` | `<Plan>` block encoding step dependencies as a Petri net | | `Transient Execution Prompt` | `<Execution>` block with step-by-step CoT paragraphs | | `Conclusion` | Final explanation and answer derived from the reasoning | | `messages` | Model-agnostic conversation format `[{role, content}, ...]` | ### Data Instance ```json { "id": "1", "Question": "Urogenital Diaphragm is made up of the following, except:", "Options": "Answer Choices:\nA. Deep transverse Perineus\nB. Perinial membrane\nC. Colle's fascia\nD. Sphincter Urethrae", "Original Reason Path": "1. Urogenital diaphragm -> Pelvic anatomy -> Muscles and fascia -> Deep perineal pouch\n2. Urogenital diaphragm -> Components of the perineum -> Muscles and connective tissue\n3. Urogenital diaphragm -> Anatomical structures -> Superficial vs. deep fascia", "Transient Plan Prompt": "<Plan>\n<Outline> Transient Step 1: Urogenital diaphragm -> Pelvic anatomy; Dependency: [] </Outline>\n...\n</Plan>", "Transient Execution Prompt": "<Execution>\n<Step> Transient Step 1: Urogenital diaphragm -> Pelvic anatomy\nThe urogenital diaphragm is part of the pelvic floor...\n</Step>\n...\n</Execution>", "Conclusion": "Explanation: Colle's fascia is part of the superficial fascia and does not contribute to the urogenital diaphragm.\nAnswer: C. Colle's fascia", "messages": [ {"role": "user", "content": "Question: \nUrogenital Diaphragm is made up of the following, except:\nAnswer Choices:\nA. ..."}, {"role": "assistant", "content": "<Think>\n...\n</Think>\nAnswer: ..."} ] } ``` ### `messages` Field The `messages` column contains the full training conversation in standard `[{role, content}]` format with no model-specific tokens. The assistant turn includes the complete DAG reasoning trace: ``` <Think> First find a reasoning path, then transform that path into an outlined plan, then execute the plan, and finally synthesize a concise conclusion. Finding Reasoning Path: 1. Entity A -> Entity B -> Entity C <Plan> <Outline> Transient Step 1: Entity A -> Entity B; Dependency: [] </Outline> ... </Plan> <Execution> <Step> Transient Step 1: Entity A -> Entity B [CoT paragraph] </Step> ... </Execution> <Conclusion> Explanation: ... Answer: X </Conclusion> </Think> Answer: [full answer text] ``` ### Usage ```python from datasets import load_dataset ds = load_dataset("Jianwen/MedVerse14k") # Use messages column for SFT training for sample in ds["train"]: messages = sample["messages"] # [{role: user, content: ...}, {role: assistant, content: ...}] ``` ### Related Resources - **Paper**: [MedVerse: Efficient and Reliable Medical Reasoning via DAG-Structured Parallel Execution](https://arxiv.org/abs/2602.07529) - **Code**: [https://github.com/aiming-lab/MedVerse](https://github.com/aiming-lab/MedVerse) ### Citation ```bibtex @article{chen2026medverse, title = {MedVerse: Efficient and Reliable Medical Reasoning via DAG-Structured Parallel Execution}, author = {Chen, Jianwen and Yang, Xinyu and Xia, Peng and Azarang, Arian and Lee, Yueh Z and Li, Gang and Zhu, Hongtu and Li, Yun and Chen, Beidi and Yao, Huaxiu}, journal = {arXiv preprint arXiv:2602.07529}, year = {2026}, url = {https://arxiv.org/abs/2602.07529} } ```