Meddies Persona VIE: 150,000 Synthetic Vietnamese Patient Personas

Demographics in, dialogue out. That is the loop most synthetic clinical data breaks at the first turn. A note reads fluently, the grammar holds, the medical vocabulary lands. Then you look at the patient underneath it and find almost nothing: no province, no occupation, no insurance status, no reason this person walked into a clinic today. The text was the easy part. The patient was the hard part, and it got skipped.

Meddies Persona VIE starts from that gap. The release is a set of 150,000 synthetic Vietnamese patient personas, built for teams that generate consultations, intake notes, triage simulations, and workflow data. The design choice is upstream. The unit is the persona, not the finished conversation. Each row is meant to give a downstream model a better patient to work from before any dialogue or note is written. Weak personas produce fluent but empty outputs. The fix is to make the patient richer before generation begins, not to polish the prose after.

A patient before a conversation

How Meddies Persona VIE is built: schema-conditioned LLM generation produces 150,000 synthetic Vietnamese patient personas as upstream context.

The argument behind the build is an ordering claim. Put the persona first, add a scenario brief, generate the draft, then filter hard before release. That order is the point. Skip the persona step and the generator has nothing to anchor to, so it produces text that sounds like a patient without being one. Give it age, background, health behavior, communication style, and social constraints, and the same generator has raw material to reason from.

Each row is machine-generated, conditioned on a schema, and written to read like a patient profile rather than a placeholder. The fields group into a few clear domains. Demographics carry age, sex, marital status, ethnicity, language, religion, province, and residence type. Social and economic context adds education, occupation, employment, household structure, insurance status, food security, digital access, and financial-literacy signals. Health and behavior context covers lifestyle factors, chronic conditions, allergies, surgeries, family history, mental-health context, environmental exposure, cultural health concepts, care-seeking behavior, and health literacy. A set of prompt-ready narrative fields then turns that structure into usable input: chief complaint, history of present illness, symptoms, social barriers, communication style, and a patient-facing narrative description. Release metadata records seeds, timestamps, model IDs, and schema versioning, so a given persona can be traced back to how it was made.

The thread running through all of it is one sentence from the card: better patient context in, better synthetic outputs out. The schema is not trying to be a medical record. It is trying to be conditioning context dense enough that a generator does not have to invent the patient from scratch.

What is dense, and what is light on purpose

A persona dataset is only as honest as its coverage map. The first practical question a reader has is which parts of the schema are filled in enough to trust as conditioning context, and which are thin. The card answers that directly rather than implying uniform completeness.

Demographics, healthcare behavior, and the narrative fields meant for the model are dense. Medication and deeper medical-history fields are lighter, and the card frames that as intent, not oversight. These personas are meant to anchor generation, not to stand in for a complete patient chart. The reading to take from this table is a usage boundary: condition on the dense domains, and treat the lighter ones as starting points a downstream step still has to fill.

The second question is spread. A dataset can be densely populated and still collapse into a handful of repeated patients. The card reports the distributions that matter for scenario-driven generation.

Ages span the full life course rather than bunching around a convenient mean. Dialect labels fall into the regional groups a Vietnamese reader would expect. Symptoms follow a long tail, which is the realistic shape, since common complaints dominate and rare ones trail off. Chronic-condition counts stay mostly low, matching a population that is not uniformly sick. Broad, varied, and still grounded is the shape you want when the goal is to stress a generator across many kinds of patient, not to model national prevalence. The card is explicit that these four views are a slice of the release and the rest is there to inspect.

What it enables, and where it stops

The intended use is narrow and stated plainly. Reach for this release when patient variation is the thing you need. 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. The persona sits at the front of the stack, and the rest of the pipeline does the work of turning it into a conversation.

Getting a first persona out takes three lines.

from datasets import load_dataset

ds = load_dataset("Meddies/meddies-persona-vie", split="train")

print(ds)
print(ds[0].keys())
print(ds[0]["demographics"])

The limits are where the card is most useful, and they are worth reading before any download. This is synthetic data, and it asks to be treated 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 the quality assurance step that filters generations before release. The dataset improves the patient you start from. It does not absolve the rest of the pipeline.

That honesty is the part worth keeping. The card does not claim the personas are real, representative, or validated against external benchmarks. It claims they are dense where density helps, varied enough to stress a generator, and light where a chart would be overreach. For a team building Vietnamese synthetic data pipelines, that is a more useful contract than a spec sheet promising completeness it cannot deliver.

The release is published under cc-by-nc-4.0 at huggingface.co/datasets/Meddies/meddies-persona-vie. Commercial use is gated behind a conversation: the card asks teams to reach out at contact@meddies.ai first. The most useful thing to send back is failure. Repetitive personas, implausible combinations, missing Vietnamese context, or downstream generations that broke in ways a QA pipeline caught are exactly the signals that sharpen the next version.