Dataset Overview

The CancerGUIDE Synthetic Patient Data dataset was released by Microsoft Research in 2025. It consists of synthetic oncology patient data generated using GPT-4.1. The associated paper can be found at 「CancerGUIDE: Cancer Guideline Understanding via Internal Disagreement Estimation」. This dataset aims to provide privacy-preserving, controllable benchmark data for evaluating and training large language models on clinical guideline adherence and reasoning tasks.

It includes synthetic oncology patient records along with corresponding treatment recommendations, generated via two methods: structured and unstructured. Specifically, there are 165 structured records and 151 unstructured records. Generated by GPT-4.1, these datasets support research into clinical reasoning and recommendation systems while avoiding the privacy risks associated with real-world patient data.

Dataset Composition

The dataset comprises two configuration files (configs), each corresponding to different generation approaches:

• synthetic_structured: Contains 165 records, generated using a structured approach (table prompt templates).
• synthetic_unstructured: Contains 151 records, generated using an unstructured approach (free-text narratives).

Each record (in JSON format) contains the following fields:

• patient_id: Unique identifier for the patient
• patient_note: Textual description of the synthesized patient's medical history
• label: Treatment plan recommended by the model

Citation

@article{CancerGUIDE,
  title = {CancerGUIDE: Cancer Guideline Understanding via Internal Disagreement Estimation},
  author = {Unell,  Alyssa and Codella,  Noel C. F. and Preston,  Sam and Argaw,  Peniel and Yim,  Wen-wai and Gero,  Zelalem and Wong,  Cliff and Jena,  Rajesh and Horvitz,  Eric and Hall,  Amanda K. and Zhong,  Ruican Rachel and Li,  Jiachen and Jain,  Shrey and Wei,  Mu and Lungren,  Matthew and Poon,  Hoifung},
  url = {https://doi.org/10.48550/arxiv.2509.07325},
  publisher = {arXiv},
  year = {2025},
}