Meddies/meddies-persona-vie

Meddies Persona VIE

Vietnamese synthetic patient personas for teams that need better patient context, not just better-looking generated text.

This is a research and simulation artifact for healthcare AI teams building Vietnamese synthetic data pipelines. It is not a record of real patients, not a prevalence reference, and not a tool for clinical decision-making. If you want to use this dataset in commercial work, please contact us at contact@meddies.ai.

Why this dataset

Most synthetic clinical data fails before the first dialogue.

Whether you are simulating doctor-patient chats, generating intake notes, or testing triage workflows, the bottleneck usually isn't the LLM, but the patient.

The note may sound fluent, but the patient behind it is thin: weak social context, weak clinical texture, weak cultural grounding. Meddies Persona VIE is built to fix that upstream.

This release contains 150,000 synthetic Vietnamese patient personas for teams generating consultations, intake notes, triage simulations, and workflow data. The unit here is the persona, not the finished conversation. Each row gives downstream generation a better patient to work from before any note or dialogue is written.

What is in this dataset

Each row is built to feel like a patient profile, not a placeholder. Records may include:

• Demographics: age, sex, marital status, ethnicity, language, religion, province, and residence type.
• Social and economic context: education, occupation, employment, household structure, insurance status, food security, digital access, and financial-literacy signals.
• Health and behavior context: lifestyle factors, chronic conditions, allergies, surgeries, family history, mental-health context, environmental exposure, cultural health concepts, care-seeking behavior, and health literacy.
• Prompt-ready narrative fields: chief complaint, history of present illness, symptoms, social barriers, communication style, and patient-facing narrative descriptions.
• Release metadata: seeds, timestamps, model IDs, and schema versioning.

The point is simple: better patient context in, better synthetic outputs out.

Coverage and diagnostics

Figure 1 answers the first practical question: which parts of the schema are dense enough to trust as conditioning context?

Demographics, healthcare behavior, and LLM-facing narrative fields are dense. Medication and deeper medical-history fields are lighter by design. These personas are meant to anchor generation, not pretend to be complete patient charts.

Figure 1. Coverage by schema domain. Some parts of the schema are strong conditioning context. Others stay intentionally light.

Figure 2 answers the next question: do the rows have enough spread to be useful?

Yes. Ages cover the full life course. Dialect labels cluster in expected regional groups such as Giọng Bắc and Giọng Nam. Symptoms follow a long tail. Chronic-condition counts stay mostly low. That is the right shape for scenario-driven generation: broad, varied, and still grounded.

Figure 2. Representative distributions for age, dialect, symptoms, and chronic-condition counts.

These four views are only a slice of the release. The rest is there for you to explore.

Where it fits in the pipeline

Use this dataset at the front of the stack.

Start with a persona. Add a scenario brief. Generate the draft. Then filter hard before release.

That order matters. Weak personas produce fluent but empty outputs. Stronger personas give downstream models better raw material: age, background, health behavior, communication style, and social constraints.

Figure 3. How Meddies Persona VIE is meant to be used: persona first, scenario next, generation after, QA before release.

Quick start

from datasets import load_dataset

# Load the dataset
ds = load_dataset("Meddies/meddies-persona-vie", split="train")
p = ds[0]

# Injecting the persona into an LLM prompt
system_prompt = f"""
Act as a Vietnamese primary care doctor. Adapt your vocabulary to this patient:

- Patient: {p['demographics']['full_name']} ({p['demographics']['gender']}, {p['demographics']['province']})
- Background: {p['socioeconomic']['occupation']} (Health literacy: {p['healthcare_behavior']['health_literacy_level']})
- Chief Complaint: "{p['llm_fields']['chief_complaint']}"
"""

print(system_prompt)

Good fits

Use this release when patient variation matters.

Good fits include synthetic doctor–patient consultations, intake and HPI note generation, triage simulations, workflow testing, and prompt stress tests across age, background, health literacy, communication style, and social barriers in Vietnamese care settings.

Limits

This is synthetic data. Treat it that way.

Do not use it to estimate prevalence, hospital volume, healthcare utilization, or national distributions. Do not use it as a clinical tool. Some demographic, socioeconomic, and clinical combinations will still be unrealistic or biased. Final output quality still depends on scenario design, symptom logic, and QA.

Feedback

Send us the failures.

The most useful feedback is concrete: repetitive personas, implausible combinations, missing Vietnamese context, clinician realism review, bias concerns, or downstream generations that break in ways your QA pipeline caught.

You can find Meddies on Hugging Face at huggingface.co/Meddies and on the web at meddies.ai.

Collaboration and sponsorship

Meddies is building Vietnamese medical data and infrastructure for safer healthcare AI.

We are a small team. Compute is still tight.

If this work matters to you—sponsorship, collaboration, clinician review, or a larger conversation about the Meddies vision—email us at contact@meddies.ai.

Citation

@misc{meddies-persona-vie,
  title={Meddies Persona VIE},
  author={MeddiesAI},
  year={2026},
  url={https://huggingface.co/datasets/Meddies/meddies-persona-vie}
}