Vietnamese hospital administration runs on situations that never reach a dataset. A ward fills past capacity during a budget freeze while a new Ministry of Health rule changes what staff are allowed to do. That is where administrative decisions get hard, and it is exactly what real clinical data cannot show. Patient records and staffing logs carry private information, so they stay locked away. The reasoning hospital administrators do every day has almost no public training material behind it.
Meddies/meddies-scenarios is our attempt to close that gap with synthetic data instead of real records. It is a collection of about 99,000 Vietnamese hospital administrative scenarios, built by the Meddies Research team and released on Hugging Face. We labeled it Domain H, for hospital administration, marking it as one slice of a larger clinical-AI effort. Each row is a written situation, in Vietnamese, that a model can read, reason about, and respond to.
Why real data could not do this job
The obvious source for administrative training data is the hospital itself. That source is also the one we cannot use. Operational records name people, facilities, and decisions, and releasing them would expose patients and staff. Translated or language-agnostic alternatives exist, but they lose the part that matters most. A budget constraint phrased in English, in a generic health system, does not carry the institutional pressure that shapes a decision inside a Vietnamese public hospital. We wanted scenarios that read like the real thing without being made of anyone's real data. Synthetic generation was the way to get both.
Building a scenario as a structured problem
The design choice underneath the whole dataset is that an administrative situation is not a paragraph. It is a problem with parts, and the parts have to be separable for a model to reason about them. So we gave every row the same skeleton, twenty-four fields in all, grouped so that each group answers one question about the situation.
First, where is this happening. A scenario is pinned to a kind of facility and a place, a district hospital or a private clinic or a tertiary-care center in a named region, then narrowed to a specific department or ward inside it. Setting comes before story on purpose. The same problem reads differently in a small district unit than in a large referral hospital, and the model should see that difference before it sees the events.
Then, what is happening. The narrative core of each row records the event that set the situation in motion, the timeline, the current state, and the resolution the staff are working toward. We split situation_triggering_event from situation_desired_outcome so cause and goal stay distinct. A model can then be asked to bridge them rather than guess at both from one blob of text.
The hardest design work went into the pressures that make these problems real. Administrative decisions are rarely hard because the right answer is unknown. They are hard because money, staffing, regulation, time, and institutional politics all push at once. We encoded those pressures as their own fields, with regulatory constraints able to reference Ministry of Health policy and political constraints reflecting the dynamics common in Vietnamese public healthcare. A scenario without constraints is a planning exercise. A scenario with conflicting constraints is the job.
The table below groups the skeleton by the question each part answers.
| Question the group answers | What it captures in the row |
|---|---|
| Where is this happening | Facility type and place (district hospital, private clinic, tertiary-care center in a named region), narrowed to a specific department or ward |
| What is happening | Narrative core: the event that set the situation in motion, the timeline, the current state, and the resolution staff are working toward, with situation_triggering_event split from situation_desired_outcome |
| What makes it hard | Pressures encoded as their own fields: money, staffing, regulation, time, institutional politics, with regulatory constraints able to reference Ministry of Health policy and political constraints reflecting Vietnamese public-healthcare dynamics |
| Where it came from | Provenance: which language model generated the row and its difficulty rating |
Each row carries its own provenance. We record which language model generated it and rate its difficulty, so the dataset can be honest about where it came from and how hard each case is meant to be.
What the release is, and what it is for
The data is gated rather than open. The card on Hugging Face is public, but downloading the files takes a Hugging Face account and acceptance of the access conditions, whether you pull through the datasets library or grab files directly. The license is Creative Commons Attribution 4.0 International, CC-BY-4.0, with no non-commercial restriction. You can use, share, and adapt the scenarios for any purpose, including commercial work, as long as you attribute the source.
We built the dataset with three uses in mind. The table below sets out each one and the property of the data it rests on.
| Intended use | What it rests on |
|---|---|
| Seed more synthetic data | A model trained or prompted on these rows learns the structure and vocabulary of Vietnamese hospital workflows without touching any real patient or staff record |
| Test cases for operations, staffing, or budget advisory systems | A spread of urgency levels and constraint types to push a system against |
| Measure administrative reasoning | The field structure makes each row a problem statement, so a model's ability to find the relevant constraints, reason toward the desired outcome, and propose next steps can be measured; the difficulty rating lets evaluation split by complexity tier rather than run as one undifferentiated pile |
The difficulty rating is what keeps the third use from collapsing into an average. A system that handles the easy tier and fails the hard one looks the same as a mediocre system on a single combined score. Splitting by tier separates the two.
Honest limits
This is synthetic data, and we are not claiming it matches the texture of real hospital operations or that any system trained on it performs well in a clinic. The scenarios were generated by language models, so they inherit whatever those models lean toward, and the difficulty ratings reflect our design intent rather than a measured outcome. What the release offers is a structured, attribution-only, Vietnamese-language starting point for administrative reasoning where there was none before. The next question is whether scenarios this shape transfer to real administrative work, and that is the thing we want to find out.
The dataset is at huggingface.co/datasets/Meddies/meddies-scenarios.