Request access to BanTMedQA

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This dataset contains health questions that real people posted in online health groups. Personal identifiers were removed from the text and the questions were lightly paraphrased, but the content is still sensitive and some images (e.g. medical reports) may show identifying details. Access is granted for research use only.

Log in or Sign Up to review the conditions and access this dataset content.

BanTMedQA

A trilingual medical question-answering dataset from Bangladesh. Each of the 196 rows is one health question that a person posted in an online health group, in three parallel versions:

  • question – Banglish: Bangla written in Latin letters, the language of the original post
  • question_bn – Bangla script
  • question_en – English

97 questions come with an image (photos of symptoms, medical reports, prescriptions, …); the other 99 are text only.

Access

This dataset is gated. Log in, fill in the short form on this page and wait for your request to be approved before the files become available.

from datasets import load_dataset
ds = load_dataset("FaiyazAbdullah114708/BanTMedQA", split="test")   # after your access request is approved

Structure

One split, test, with 196 rows. Posts that appeared more than once in the source data were removed, keeping one copy.

column description
id row number (0–195)
image the attached image; null for text-only rows
image_id image identifier; null for text-only rows
data_type image-text (97) or text (99)
question, question_bn, question_en the question in Banglish, Bangla and English
answer reference answer; null when there is none (71 rows)
age, height, weight, sex patient details mentioned in the post; null when not stated
medical_category e.g. Dermatology/Skin (49), Gynecology/Obstetrics (38), Pediatrics/Child Health (16), …
question_type Story-based (111) or Direct (85)
additional Symptom-based (100), Direct (37), Narrative (36), Report Interpretation (23)

Privacy processing

  • Questions: names, usernames, @mentions, phone numbers, URLs, hashtags, group names, clinic and hospital names, small-area addresses and the year of dates were removed. All three versions were then lightly paraphrased the same way so that a post cannot be found by exact text search, while keeping the meaning, numbers, doses and medicine names. Every row was checked against its original for meaning changes, leftover identifiers and disagreement between the three versions, and corrected where needed. City and district names are kept; area names are kept only when the question is about services in that area.
  • Answers: doctors' names were replaced by their role (e.g. "a senior dermatologist"); the rest of each answer is unchanged.
  • Images: not modified. Some (especially the 23 report-interpretation questions) may show names or other identifying details.

Intended use and limitations

For research on medical question answering in Bangla, Banglish (code-mixed, transliterated) and English, including with images.

Downloads last month
5