Medical Dialogue · OpenMed
MedDialog
A dataset of 251,731 patient-doctor dialogue pairs for training and evaluating clinical conversation models.
What MedDialog is
MedDialog is a large-scale clinical dialogue dataset developed by OpenMed, consisting of 251,731 English conversation pairs between patients and physicians. The collection is designed for natural language processing tasks, specifically text generation and question answering, where models must learn to generate appropriate, empathetic, and medically sound responses to clinical enquiries.
The dataset addresses general medicine and clinical communication, encompassing patient concerns regarding symptoms, diagnostic reasoning, medications, clinical procedures, triage, and general health topics. It provides researchers with curated training and validation pairs to benchmark conversational agents against ground-truth physician responses.
How it was made
The preparation pipeline combined description and patient query fields into a single patient message, filtered out uninformative interactions with fewer than five patient words or ten physician words, and excluded responses consisting purely of referrals without medical guidance. All messages were capped at a maximum length of 2,500 characters.
The resulting 251,731 entries are organised into a training split of 226,557 examples and a validation split of 25,174 examples via a 90/10 random partition. Each record comprises three text fields: the patient message describing symptoms or questions, the ground-truth doctor response, and a dialogue context field reserved for multi-turn interactions, which is left empty in the current release.
Key facts
- Total examples
- 251,731
- Train split
- 226,557
- Validation split
- 25,174
- Language
- English
- Domain
- Clinical and general medicine
- Source dataset
- ruslanmv/ai-medical-chatbot
- Format
- JSON
- Modality
- Text
What you can use it for
- Clinical response generation
- Training language models to generate empathetic, relevant, and medically appropriate dialogue responses to patient queries.
- Medical triage and symptom assessment
- Developing automated dialogue systems that process patient-reported symptoms and provide initial health advice or recommend professional consultation.
- Reinforcement learning with human or rule-based feedback
- Serving as reference data within reinforcement learning environments to score outputs on response quality, empathy, medical accuracy, safety disclaimers, and fluency.
- Dialogue system evaluation
- Benchmarking general and clinical language models against physician responses across a wide variety of medical conditions and questions.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from OpenMed/MedDialog at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 226,557 rows of default/train
dialogue_context
string_label- 100%
doctor_response
string_text- min
- 41
- median
- 481
- mean
- 545.97
- max
- 2,500
patient_message
string_text- min
- 22
- median
- 414
- mean
- 492.45
- max
- 2,500
Strengths and limits
Good at
- Filters out low-content messages and simple referral redirects to ensure responses contain substantive medical guidance.
- Supplies pre-defined training and validation splits using a fixed random seed to facilitate reproducible experimentation.
Watch out for
- Entries reflect online medical question-and-answer interactions, which may vary in quality, contain inaccuracies, and diverge from formal clinical best practice.
- Interactions are limited to single-turn exchanges because the dialogue context field contains no text in the current version.
- The data is intended strictly for research development and cannot be used as a substitute for professional clinical advice.
Questions
Can I look inside MedDialog without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is MedDialog under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
What fields are contained in each dataset record?
Each entry contains three fields: patient_message (the patient enquiry or symptom description), doctor_response (the reference clinician response), and dialogue_context (a field reserved for multi-turn conversation context, currently empty).
Does this dataset support multi-turn conversation training?
In the current release, all examples are single-turn question-and-answer pairs, and the dialogue_context field remains empty.
What filtering was applied to the source data?
The developers merged separate patient description fields, removed short messages under five words for patients or ten words for doctors, excluded entries where doctors only offered a referral without substantive content, and truncated messages to 2,500 characters.
How are the training and validation sets divided?
The dataset is partitioned into a 90% training split of 226,557 examples and a 10% validation split of 25,174 examples using a fixed random seed of 42.
How to cite it
@dataset{openmed_meddialog_2026, title={MedDialog: A Medical Dialogue Dataset for Clinical Response Generation}, author={OpenMed}, year={2026}, publisher={Hugging Face}, url={https://huggingface.co/datasets/OpenMed/MedDialog}}
Written from MedDialog's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.