Verox132/Medical_Conversational_Dataset

Synthetic Hospital-Robot Conversational Dataset

A fully synthetic dataset of multi-turn conversations between a patient (or visitor) and a hospital service robot, generated for training and evaluating conversational AI models on human-robot interaction (HRI) in a medical setting.

This is not real patient data. Every name, patient ID, room number, and appointment is randomly generated at build time — nothing in this dataset was collected from an actual hospital, patient, or interaction. It is also not a source of medical advice or clinical guidance; several of the conversations exist specifically to show the robot declining to give medical advice and escalating to human staff instead.

Dataset Summary

Conversations are built from hand-authored templates covering eight behavioral categories of hospital-robot interaction:

The current release (Test_Dataset_9) contains 61,350 conversations generated from 128 distinct templates, split 90/10 into train and validation.

Dataset Structure

Each line of the .jsonl files is one JSON object with a single field, messages: a list of chat turns in strict user/assistant(/tool) alternation, in the widely-used chat-format style (compatible with ChatML-style and OpenAI-style chat templates).

Two file variants are provided, built from the same underlying conversations:

train.jsonl / validation.jsonl (default)

The generator's native format. A tool call is a single object on the assistant message; a tool's response is a plain JSON object.

{
  "messages": [
    {"role": "user", "content": "Could you walk Helga Schulz over to 1015 for the appointment?"},
    {"role": "assistant", "content": "Of course, on our way to 1015 now."},
    {"role": "user", "content": "Actually, stop — 1015 just got closed off, go to the Sunroom instead."},
    {
      "role": "assistant",
      "content": "No problem, updating the route to the Sunroom.",
      "tool_call": {"name": "navigate_to", "arguments": {"room": "the Sunroom"}}
    },
    {"role": "tool", "name": "navigate_to", "content": {"status": "arrived", "location": "the Sunroom"}},
    {"role": "assistant", "content": "Here at the Sunroom now, no trouble redirecting."}
  ]
}

train_qwen_61k.jsonl / validation_qwen_61k.jsonl (qwen config)

The same conversations reshaped into the OpenAI/Qwen function-calling format: tool_calls is a list with an id, and the matching tool response carries tool_call_id with its content JSON-stringified. Use this variant if you're fine-tuning a model that expects that convention.

{
  "messages": [
    {"role": "user", "content": "..."},
    {
      "role": "assistant",
      "content": "No problem, updating the route to the Sunroom.",
      "tool_calls": [{"id": "call_45cc3a5b", "type": "function", "function": {"name": "navigate_to", "arguments": {"room": "the Sunroom"}}}]
    },
    {"role": "tool", "name": "navigate_to", "tool_call_id": "call_45cc3a5b", "content": "{\"status\": \"arrived\", \"location\": \"the Sunroom\"}"},
    {"role": "assistant", "content": "..."}
  ]
}

Load either variant with the datasets library:

from datasets import load_dataset

ds = load_dataset("Verox132/Medical_Conversational_Dataset", "default")
# or: load_dataset("Verox132/Medical_Conversational_Dataset", "qwen")

Source Data / Generation Methodology

Every conversation is produced by a template-driven generator: a scenario defines the entities involved (patient, appointment, weather, etc.) and any constraints between them, a language schema supplies pools of phrasing for each turn, and a JSON template assembles them into a full conversation. Every rendered conversation is validated (structure, resolved placeholders, matched tool calls, pronoun/gender consistency) before being kept, and consecutive same-role turns are folded so the result is always strict user/assistant alternation.

The generator itself is open source: [github.com/WoBen07/Synthetic_Conversational_Dataset_Generator](https://github.com/WoBen07/Synthetic_Conversational_Dataset_Generator)

This project was started as part of a Bachelor's thesis at the University of Augsburg, Germany, developed by Benedict Wohlgemuth under the guidance of Dr. Matthias Kraus. The dataset is under active development — more categories, entrypoints, and dataset versions will be added over time under new subfolders in this repo.

Evaluation

An evaluation of models on this dataset is planned and will follow soon. Results, the evaluation setup, and the metrics used will be added to this section once they are available.

Limitations

• —Synthetic, not real-world data. Language patterns reflect the authors' template design, not observed real hospital interactions.
• —English only, and reflects a Western/German-influenced naming and cultural context (patient and visitor names are drawn from a German name pool).
• —Template-based diversity. Phrasing is drawn from finite pools of hand-written options per template, so exact wording repeats across the dataset far more than in naturally collected dialogue.
• —Not clinically validated. Nothing in this dataset should be used as a source of medical advice, and models trained on it should not be deployed to give medical advice either.

License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

Copyright © 2026 Benedict Wohlgemuth, University of Augsburg.

You are free to share and adapt the dataset, including for commercial purposes, as long as you give appropriate credit to the author and the University of Augsburg, provide a link to the license, and indicate if changes were made. Please use the citation below.

Citation

If you use this dataset, please cite the source repository:

Wohlgemuth, B. (2026). Synthetic Conversational Dataset Generator
[Software and dataset]. University of Augsburg.
https://github.com/WoBen07/Synthetic_Conversational_Dataset_Generator