Dialt/simulated-patients

▶ Run these cases on Dialt: get an API key

Dialt develops custom simulated characters for your team, product or training programme. For inquiries, email hello@dialt.com.

Dialt Simulated Patients

Six fictional adults for non-emergency history-taking and communication practice. Each scenario supplies a patient's history, concerns and conversational behavior, with a separate interviewer brief. Use the text with your own conversational model or build a voice-based exercise on Dialt. No clinical or educator validation is claimed.

Choose an exercise

Simulated patient	Interview task	What makes it different
Martin	Explore treatment experience and practical barriers	Embarrassed about missed tablets; understanding an explanation need not resolve his concern
Robin	Explore sleep and stress with permission and sensitivity	Guarded about private family pressures; privacy is compatible with truthful answers
Alex	Establish a clear symptom chronology	Anxious and prone to tangents; structure should help without dismissing concern
Priya	Understand why a previously discussed exercise plan was not followed	Embarrassed about a plan that does not fit daily life; willingness is different from a feasible commitment
Daniel	Understand an unresolved follow-up concern	Frustrated by delay; an apology is different from a confirmed next step
Evelyn	Establish accessible communication and elicit concerns	Initially minimizes difficulty following speech; clarification preserves agency

These are situations with facts, motivations and conditional responses. Demographics do not determine behavior. The characters answer direct factual questions truthfully, do not prescribe and do not invent unknown medical history. They identify themselves as simulated patients when directly asked.

What is included

Each row has a stable id, case version, title, review status, public_brief, fixed facts, baseline_instructions, language, an original voice_brief, a short delivery_prompt, assistant_interruptions, source attribution and five practice-quality criteria. The source field preserves pins for Martin's adopted synthetic facts; the other cases are original fictional authoring. No customer records, real patient conversations or real speaker recordings seed those new cases.

The practice split contains all six cases with their full briefs and practice criteria. Keep separate, unexposed cases for assessment. Held-out probes and assessment results are not included.

Load and use the scenarios

Install the loading library with pip install datasets. While this repository is private, authenticate with hf auth login using a Hugging Face token with repository access.

from datasets import load_dataset

patients = load_dataset("Dialt/simulated-patients", split="practice", token=True)
assert len(patients) == 6
patient = next(row for row in patients if row["id"] == "patient-martin-01")
patient_instructions = patient["facts"] + "\n\n" + patient["baseline_instructions"]
interviewer_context = patient["public_brief"]

To export from the Dialt source repository, run:

uv run python voices/scripts/export_patient_pack.py --output /tmp/patient-pack-review

The exporter writes practice.jsonl, this card, licenses and a SHA-256 manifest into a new directory. Load that file with load_dataset("json", data_files="/tmp/patient-pack-review/practice.jsonl", split="train").

Give only public_brief to the interviewer. Give facts and baseline instructions to the patient runtime, and facts and criteria to the grader. Do not give the grader's expected answers to the interviewer. Instruction hierarchy and safety remain the integrator's responsibility. A downloadable scenario can be used with an independently supplied conversational model; this release contains neither a standalone simulator nor model weights.

Use through Dialt

Dataset access does not grant hosted voice access. To run a scenario in your organization, create your own saved Design using an available speaker base. The download includes voice briefs, not voice identities or recordings.

To build an independent Design, select an available speaker base in Voice Design, set instructions to facts plus baseline instructions, conditioning_prompt to the delivery prompt, assistant_interruptions to the supplied permission and brain to smart, then save and publish a named immutable Design. A stock base exercises the behavior without reproducing the bespoke identity described by voice_brief. The brief is an authoring input; the download does not contain the voice or a selectable provider handle.

The Voice Design API guide documents the same authoring operations. Once published in your organization, the ordinary session mode is:

{"kind": "dialt", "design": "YOUR_PUBLISHED_DESIGN_UUID"}

The same Design ID works in Playground and in Evals as target.design when checking the patient, or simulator.design when testing an interviewer agent. Patient-quality criteria do not assess clinician competence. See the Evals guide for cases and runs.

Limitations

These are incomplete fictional histories for communication practice. They are unsuitable for diagnosis, prescribing, emergency triage or formal clinical assessment. Unknown clinical details remain unknown; absence from a case is not a negative clinical finding. There has been no educator review, clinical validation, accreditation or demonstration of educational effectiveness.

The supplied instructions are retained baselines, with no demonstrated optimization win. Engineering checks cover a small set of conversations with these fixed characters; they do not establish generalization to new patients, dialect authenticity or suitability for a training curriculum. The exposed criteria need independent validation for your use case.

License and provenance

Scenario text, this card and JSONL exports are CC BY 4.0. Commercial reuse, adaptation and model training are permitted under that license, with attribution and indication of changes. See LICENSE-SCENARIOS for attribution and scope. The export script and code examples are MIT, under LICENSE-CODE. These grants do not include hosted services, trademarks, voice identities, recordings or model weights.

Martin adopts essential hypertension and the names/doses of two medicines from MITRE Synthea synthetic sample data under its download policy. His new name, demographics, circumstances and portrayal are original fictional additions; the source pins are in his row. The other five cases are original fictional scenarios. No real person's likeness is used. Public availability can lead to copying or training overlap, so public cases must never be represented as unseen assessments.