The MedExAgent data contains synthetic doctor-patient conversations and patient case tables derived from DDxPlus and PMC-Patients-v2. It is intended for research on medical agents in simulated settings. It is not medical advice and must not be used for clinical care, diagnosis, treatment decisions, emergency triage, patient management, or any other real-world medical decision-making. By requesting access, you agree to use the data for non-commercial research only and to comply with the CC BY-NC-SA 4.0 license and the licenses of the source datasets.

MedExAgent Data

Data used to train and evaluate MedExAgent-8B, from the paper "MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments" (NeurIPS 2026).

Paper: https://openreview.net/forum?id=olngQeJA7J (arXiv: https://arxiv.org/abs/2605.07058)
Code: https://github.com/EndlessCG/medexagent
Model: https://huggingface.co/medagent/MedExAgent-8B

Files

File	Rows	Used for
`sft/train.jsonl`	5,984	SFT stage: synthetic doctor-patient conversations
`rl/train_patients.csv`	19,984	RL stage: patient case tables (demographics, history, symptoms, recorded exams and findings)
`rl/train_verl.parquet`	19,984	The same RL cases in the verl format used for DAPO training, with per-case noise assignments
`eval/ddxplus_val.jsonl`, `eval/pmc_val.jsonl`	2,497 / 2,499	Validation
`eval/ddxplus_test.jsonl`, `eval/pmc_test.jsonl`	2,497 / 2,498	Test, clean
`eval/ddxplus_test_all_noise.jsonl`, `eval/pmc_test_all_noise.jsonl`	2,499 / 2,498	Test, with patient and exam noise

Training cases come from DDxPlus and PMC-Patients-v2 in roughly equal parts. The paper reports results on 1,000 cases from each test file.

Conversation format

Each line of the .jsonl files is one patient:

patient_id, diagnosis, demographics, medical_history, self_reported_symptoms
patient_persona: personality, language proficiency, and medical history recall, each with a key, a label, and the instruction given to the patient simulator
conversation: a list of messages with roles system, assistant (doctor), user (patient), and tool (exam results). Exams are ordered as tool calls
noise_metadata: the patient and exam noise applied to this conversation and where it was applied

Conversations follow the five Calgary-Cambridge stages. They were generated with gpt-4o-mini as both doctor and patient, with patient noise applied to 30% of patients and exam noise to 10% of eligible exam results.

License and sources

This dataset is released under CC BY-NC-SA 4.0. It is derived from:

DDxPlus (Fansi Tchango et al., 2022), CC BY 4.0
PMC-Patients-v2 (Zhao et al., 2023), CC BY-NC-SA 4.0
The patient persona axes follow PatientSim (Kyung et al., 2025), CC BY-NC-ND 4.0

The data contains no exam price values. Exam costs used in training are computed by the code from cost and discomfort tiers.

Intended use

Research on medical agents in simulated, non-clinical settings only. The conversations are synthetic and may contain incorrect medical information. The data must not be used for clinical care or any other real-world medical decision-making.

Citation

@misc{gao2026medexagenttrainingllmagents,
      title={MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments},
      author={Yicheng Gao and Xiaolin Zhou and Yahan Li and Yue Zhao and Ruishan Liu},
      year={2026},
      eprint={2605.07058},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.07058},
}

Downloads last month: 52

Total file size:

282 MB

Paper for medagent/MedExAgent-Data

MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments

Paper • 2605.07058 • Published May 8 • 2