ASCRIBE
Collection
Datasets released as a part of the contributions of "ASCRIBE: Atomic and Significance-Based Reasoning for Thai Clinical SOAP Note Generation" • 3 items • Updated
Request access to ThaiClinicBench
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
ThaiClinicBench contains de-identified real clinical encounters collected with informed consent. It is released under the ThaiClinicBench Research Data Use Agreement. Please read the full agreement below before requesting access.
ThaiClinicBench Research Data Use Agreement
These terms govern access to and use of the ThaiClinicBench dataset (the "Dataset"), released by Looloo Technology Co., Ltd. (the "Dataset Providers") with the paper ASCRIBE: Atomic and Significance-Based Reasoning for Thai Clinical SOAP Note Generation (WiNLP @ EMNLP 2026).
1. Acceptance of Terms
Requesting access to or using the Dataset means you agree to these terms. If you disagree with any part, do not use the Dataset.
2. Permitted Use
- The Dataset is intended solely for academic, research and educational purposes in clinical language processing and medical-scribe systems, for evaluation only.
- The Dataset is a held-out benchmark. Do not use it to train or fine-tune models, or for prompt or hyperparameter selection in a way that compromises its use as a held-out test set.
- Commercial use of the Dataset without prior written permission from the Dataset Providers is forbidden.
- You must follow all applicable laws, regulations and research ethics, including data privacy and protection standards.
3. Data Protection and Privacy
- The Dataset contains de-identified real clinical encounters. You must keep the Dataset confidential.
- Do not attempt to identify, or re-identify, any patient, physician, clinic or other individual or institution described in the Dataset, whether from the Dataset alone or in combination with other information.
- If you become aware that any individual could be identified from the Dataset, stop using the affected data and notify the Dataset Providers promptly (see Section 12).
- Store the Dataset securely, apply reasonable safeguards against unauthorized access, and limit access to individuals who have themselves accepted these terms.
4. No Clinical Use
Do not use the Dataset, or any model or system developed with it, to make or support diagnosis, treatment or any other clinical decision about real patients.
5. Redistribution
- Redistribution of the Dataset, or any portion of it, is not allowed. This includes sharing, publishing, re-hosting, selling or sublicensing it. Anyone else who wishes to use the Dataset must request access from the Dataset Providers.
- Sharing derived data must respect the privacy and confidentiality terms set out above.
6. Attribution
- Cite the ASCRIBE paper in any publication or public work that uses the Dataset.
- You may not claim ownership of, or exclusive rights over, the Dataset or its derivatives.
7. Access Revocation and Data Removal
- Violation of these terms may result in the termination of your access to the Dataset.
- On termination, or on request from the Dataset Providers (including in response to a data-removal request from a participant), you must delete all copies of the Dataset, or of the affected portion, in your possession.
8. Disclaimer
The Dataset is provided "as is", without warranty of any kind, either express or implied, including but not limited to the accuracy or completeness of the data.
9. Limitation of Liability
Under no circumstances will the Dataset Providers be liable for any claims or damages resulting from your use of the Dataset.
10. Amendments
These terms may be updated at any time. Continued use of the Dataset signifies acceptance of the updated terms.
11. Governing Law
These terms are governed by the laws of Thailand, excluding its conflict-of-law rules.
12. Contact
Questions, privacy concerns, data-removal requests and incident reports should be sent to: contact@looloohealth.com
Consent: Requesting access to or using the Dataset signifies your acknowledgment of and agreement to these terms.
Requests are reviewed manually. Please use your institutional email and describe your intended use.
Log in or Sign Up to review the conditions and access this dataset content.
ThaiClinicBench
ThaiClinicBench is a benchmark of 44 de-identified, real outpatient (OPD) encounters from a medical clinic in Thailand. Each encounter pairs a doctor–patient conversation transcript with the physician-written SOAP note. To our knowledge, it is the first Thai medical-scribe benchmark released for research.
It was released with the paper ASCRIBE: Atomic and Significance-Based Reasoning for Thai Clinical SOAP Note Generation (WiNLP @ EMNLP 2026), where it serves as an out-of-distribution test set.
Dataset Summary
| Encounters | 44 (test only) |
| Language | Thai (colloquial, with dialectal medical terms) |
| Transcript length | 703 words (2.89k characters) on average |
| Transcripts | Produced manually by human annotators |
| Notes | Physician-written SOAP notes |
| Audio | Not included |
Data Fields
Each record is one encounter.
| Field | Type | Description |
|---|---|---|
id |
int | Public record id (1–44), assigned at random. |
transcript |
list of string | The doctor–patient conversation, in order. Each element is a segment of consecutive speech and can contain turns from both speakers. Speaker labels are not included. |
soap_note |
object | The physician-written SOAP note, as four lists of strings: history_of_present_illness (Subjective), physical_exam and results (Objective), and assessment_and_plan. A section with no content is an empty list. |
atomic_facts |
list of object | Model-extracted reference atomic facts for the encounter (see the paper). Each has atomic_bullet (one clinical fact, in Thai), medical_significance (one of Required, Diagnosis-impacting, Treatment-impacting, Good for follow-up, For future, Not needed) and evidence_note (a short English justification). |
The data only contains the test split.
Collection and De-identification
Encounters were collected with signed informed consent from all participating patients and physicians, who were told that their clinical dialogue data would be de-identified and used for AI research and dataset development. Participation was voluntary.
All transcripts and notes were manually de-identified before use. Direct identifiers and potentially identifying context (names, contact information, locations, dates, institutional identifiers and other encounter-specific details) were removed or replaced with synthetic placeholders, following the HIPAA Safe Harbor method and Thailand's Personal Data Protection Act (PDPA). See App. H of the paper.
Uses
Use of ThaiClinicBench is governed by the ThaiClinicBench Research Data Use Agreement. The examples below are illustrations, not a complete list. Where they differ from the agreement, the agreement applies.
Examples of intended use
- Evaluating a system that generates SOAP notes from Thai doctor–patient conversations, and reporting its scores in a paper (§2).
- Studying how colloquial and dialectal Thai affects clinical note generation (§2).
- Analyzing errors in generated notes, such as omitted or unsupported findings (§2).
- Teaching a clinical NLP course in which each student requests access themselves (§2, §3).
Examples of prohibited use
- Training or fine-tuning a model on the encounters (§2).
- Tuning prompts or hyperparameters on the benchmark, then reporting the tuned score as a held-out result (§2).
- Using benchmark results to market a commercial product without written permission (§2).
- Sharing the files with a colleague who has not requested access, or uploading them to a public repository. (§3, §5).
- Trying to identify a patient, physician or clinic described in an encounter, including by combining it with other information (§3).
- Using a system developed with the dataset to suggest diagnoses or treatments for real patients (§4).
License
Released under the ThaiClinicBench Research Data Use Agreement, version 1.0. See LICENSE. Access is granted on request after manual review.
Limitations
The benchmark is small (44 encounters) and comes from a single clinic, so it may not represent the full range of Thai clinical settings, specialties or dialects. Placeholders introduced during de-identification may make some passages read less naturally.
Contact
For questions, privacy concerns or data-removal requests, contact contact@looloohealth.com.
Citation
@misc{kalavantavanich2026ascribe,
title = {ASCRIBE: Atomic and Significance-Based Reasoning for Thai Clinical SOAP Note Generation},
author = {Tarm Kalavantavanich and Teerawut Ponarchar and Pattaramanee Arsomngern and Jenta Wonglertsakul and Watcharakorn Chuthong and Chiraphat Boonnag and Knot Pipatsrisawat and Titipat Achakulvisut},
year = {2026},
eprint = {2610.01234},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2610.01234}
}
- Downloads last month
- 43