🩺 BanglaMedBench

A Bangla (বাংলা) Medical Question Answering Benchmark for Evaluating Large Language Models

500 medical MCQs · 39 specialties · Basic + Clinical · Bangla ↔ English parallel

BanglaMedBench-500 is a balanced benchmark for testing how well large language models (LLMs) understand and answer medical questions in Bangla (Bengali), the 7th most spoken language in the world, with over 270 million speakers. It is built from widely used English medical benchmarks (MedQA, MedMCQA, PubMedQA, MMLU), translated into Bangla with Gemma 4, and organised by medical specialty so you can see where a model is strong or weak.

🇧🇩 সংক্ষেপে: এটি বাংলা ভাষায় চিকিৎসা-বিষয়ক ৫০০টি বহুনির্বাচনী প্রশ্নের একটি বেঞ্চমার্ক। এর মাধ্যমে যেকোনো ভাষা-মডেল (LLM) বাংলা মেডিকেল প্রশ্নের উত্তর কতটা ভালো দিতে পারে, তা যাচাই করা যায়। প্রশ্নগুলো ৩৯টি চিকিৎসা-শাখায় ভাগ করা এবং প্রতিটির ইংরেজি ও বাংলা দুটি সংস্করণই আছে।

---

✨ Why BanglaMedBench-500?

• Low-resource gap. Bangla has very few medical evaluation sets. Most medical LLM benchmarks are English or Chinese only.
• Balanced by design. 13 questions per specialty (except one), so no single field dominates the score.
• Basic vs Clinical split. Separate scores for basic medical science (anatomy, physiology, pharmacology...) and clinical practice (cardiology, paediatrics, emergency medicine...).
• Parallel English + Bangla. Evaluate the same question in both languages and measure the cross-lingual gap.
• Quality signal. Every row carries a translation confidence score, so you can run a high-confidence-only evaluation.
• Small and fast. 500 questions run in minutes, so it is cheap to evaluate many models.

🇧🇩 কেন এই বেঞ্চমার্ক? বাংলায় চিকিৎসা-বিষয়ক মডেল-মূল্যায়নের ডেটাসেট খুবই কম। এখানে প্রতিটি শাখা থেকে সমান সংখ্যক প্রশ্ন রাখা হয়েছে, মৌলিক (Basic) ও ক্লিনিক্যাল (Clinical) আলাদা করা হয়েছে, এবং ইংরেজি-বাংলা তুলনা করার সুযোগ আছে।

---

📊 At a glance

Total questions	500
Format	Multiple choice (480 with 4 options, 20 with 3 options: yes/no/maybe)
Languages	Bangla (*_bn fields) + original English
Groups	Basic: 253 · Clinical: 247
Specialties	39 (13 questions each; Basic Psychiatry has 6) (confidence: classification confidence number)
Translation	Gemma 4 (machine translation)
License	Apache-2.0

Where the questions come from

Source	Questions	Original license
MedQA (USMLE-style)	286	MIT
MedMCQA (AIIMS / NEET-PG)	125	Apache-2.0
PubMedQA	20	MIT
MMLU medical subsets	69	MIT

MMLU subsets used: college_medicine (17), clinical_knowledge (11), virology (10), professional_medicine (10), anatomy (10), college_biology (6), medical_genetics (5).

Questions per specialty (click to expand)

13 questions each for: Allergy & Immunology, Anaesthesiology, Anatomy, Basic ENT, Basic Neuroscience, Basic Obstetrics & Gynaecology, Basic Ophthalmology, Basic Orthopaedics, Basic Paediatrics, Basic Surgery, Biochemistry, Cardiology & Vascular Medicine, Dermatology, ENT & Oral Health, Endocrinology & Metabolism, Epidemiology & Public Health, Forensic Medicine, Gastroenterology & Hepatology, Genetics & Molecular Biology, Geriatrics & Palliative Care, Infectious Diseases & Tropical Medicine, Microbiology & Immunology, Nephrology & Urology, Neurology, Obstetrics & Gynaecology, Oncology & Haematology, Ophthalmology, Other Basic Sciences, Other Clinical, Paediatrics & Neonatology, Pathology, Pharmacology, Physiology, Psychiatry & Mental Health, Radiology & Imaging, Respiratory Medicine, Rheumatology & Musculoskeletal, Trauma & Emergency Medicine.

6 questions: Basic Psychiatry.

---

🚀 Quick start

pip install datasets

from datasets import load_dataset

ds = load_dataset("Arittra/BanglaMedBench-500", split="test")
print(ds[0]["question_bn"])
print(ds[0]["options_bn"])
print(ds[0]["answer_bn"])

Evaluate any model

import re
from datasets import load_dataset

ds = load_dataset("Arittra/BanglaMedBench", split="test")
LETTERS = "ABCD"

def build_prompt(row, lang="bn"):
    q = row["question_bn"] if lang == "bn" else row["question"]
    opts = row["options_bn"] if lang == "bn" else row["options"]
    body = q + "\n" + "\n".join(f"{LETTERS[i]}. {o}" for i, o in enumerate(opts))
    return body + ("\n\nউত্তর (শুধু অক্ষর):" if lang == "bn" else "\n\nAnswer (letter only):")

def correct_letter(row):
    if row["answer"] not in row["options"]:
        return None                      # skips the 20 broken PubMedQA rows until you fix them
    return LETTERS[row["options"].index(row["answer"])]

def extract_letter(text):
    m = re.search(r"\b([A-D])\b", text.upper())   # finds "B" even in "The answer is B."
    return m.group(1) if m else None

def evaluate(generate_fn, lang="bn", limit=None):
    hits = total = 0
    for i, row in enumerate(ds):
        if limit and i >= limit: break
        gold = correct_letter(row)
        if gold is None: continue
        pred = extract_letter(generate_fn(build_prompt(row, lang)))
        hits += pred == gold
        total += 1
    return hits / total

# ---------- OPTION 1: a model on your own computer (Hugging Face transformers) ----------
from transformers import pipeline
pipe = pipeline("text-generation", model="YOUR_MODEL_ID_HERE", device_map="auto")
# e.g. a Gemma model you already downloaded; must be a chat/instruct model

def my_model_generate(prompt):
    out = pipe([{"role": "user", "content": prompt}], max_new_tokens=10, do_sample=False)
    return out[0]["generated_text"][-1]["content"]

# ---------- OPTION 2: Ollama (if you run models with Ollama) ----------
# import requests
# def my_model_generate(prompt):
#     r = requests.post("http://localhost:11434/api/chat", json={
#         "model": "YOUR_OLLAMA_MODEL_NAME",
#         "messages": [{"role": "user", "content": prompt}],
#         "stream": False})
#     return r.json()["message"]["content"]

# ---------- OPTION 3: a paid API (example: Anthropic) ----------
# import anthropic
# client = anthropic.Anthropic(api_key="YOUR_API_KEY")
# def my_model_generate(prompt):
#     msg = client.messages.create(model="claude-sonnet-5-5", max_tokens=10,
#                                  messages=[{"role": "user", "content": prompt}])
#     return msg.content[0].text

print("Test run, 20 questions:", evaluate(my_model_generate, "bn", limit=20))
# print("Bangla:",  evaluate(my_model_generate, "bn"))
# print("English:", evaluate(my_model_generate, "en"))

🗂️ Data fields

Field	Description
id	Unique ID, e.g. BN-BENCH-0001
question / question_bn	Question in English / Bangla
options / options_bn	Answer choices in English / Bangla (same order)
answer / answer_bn	Correct answer text in English / Bangla
source	Original benchmark (MedQA, MedMCQA, PubMedQA, MMLU-subset)
category	One of 39 medical specialties
group	Basic or Clinical
confidence	Translation confidence score (0 to 1)

🇧🇩 ফিল্ড-পরিচিতি: question_bn, options_bn, answer_bn হলো বাংলা সংস্করণ; category হলো চিকিৎসা-শাখা; group মানে মৌলিক (Basic) নাকি ক্লিনিক্যাল (Clinical); confidence হলো অনুবাদের মান-স্কোর।

---

🏆 Leaderboard

Results will be added as models are evaluated. Want your model listed? Open a discussion on this repo with your scores and evaluation setup.

Model	Bangla acc.	English acc.	Gap	Notes
(your model here)	–	–	–	–

---

🛠️ How it was built

1. Selection. 500 questions were sampled from MedQA, MedMCQA, PubMedQA and MMLU medical subsets.
2. Classification. Each question was assigned to one of 39 specialties and to a Basic or Clinical group, with balancing across specialties.
3. Translation. Question, options and answer were machine-translated into Bangla with Gemma 4. A confidence score is stored per row. Common English medical terms are kept in brackets next to the Bangla term to reduce ambiguity.
4. Quality checks. Rows were checked for missing translations, option/answer consistency and duplicates.

🇧🇩 কীভাবে তৈরি: ইংরেজি বেঞ্চমার্ক থেকে ৫০০টি প্রশ্ন বেছে নিয়ে শাখা অনুযায়ী ভাগ করা হয়েছে, তারপর Gemma 4 দিয়ে বাংলায় অনুবাদ করা হয়েছে। পরিভাষা বোঝার সুবিধার জন্য অনেক জায়গায় ইংরেজি শব্দ বন্ধনীতে রাখা হয়েছে।

---

⚠️ Limitations and responsible use

• Evaluation only. Please do not train on this benchmark. Doing so would make results meaningless.
• Possible contamination. The questions come from public benchmarks that may already appear in some models' training data. The companion training dataset (Arittra/bangla_medical_qa_cot_and_non_cot) is built from sources that share upstream material with this benchmark; a few near-duplicates were checked and removed. Treat very high scores with care.
• Machine translation. Translations were not fully reviewed by medical professionals. Terminology, numbers or units may contain errors. Use the confidence field to filter.
• Origin bias. Questions come largely from US and Indian exams; they do not fully reflect Bangladeshi disease patterns, drug brands or national guidelines.
• Small size. 500 questions give a useful signal, but per-specialty results (13 questions each) have wide error bars. Do not over-interpret small differences.
• Not medical advice. This dataset is for research and model evaluation only, never for clinical decisions.

---

📚 Citation

If you use BanglaMedBench-500, please cite it and the original benchmarks.

@misc{banglamedbench500,
  title        = {BanglaMedBench-500: A Bangla Medical Question Answering Benchmark},
  author       = {Arittra},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Arittra/BanglaMedBench-500}}
}

Original sources: MedQA (Jin et al., 2020), MedMCQA (Pal et al., 2022), PubMedQA (Jin et al., 2019), and MMLU (Hendrycks et al., 2021). Please verify the exact citation details on each source's page.

---

🔗 Related resources

• 📦 Training data: Bangla Medical QA (CoT and Non-CoT)
• 💻 Code: [add your GitHub link]

📄 License and attribution

Released under Apache-2.0. This benchmark is a derivative of MedQA (MIT), MedMCQA (Apache-2.0), PubMedQA (MIT) and MMLU (MIT); original copyright notices and attribution remain with their authors. Changes made: sampling, specialty classification, machine translation to Bangla, and consistency checks.

---

If this benchmark helps your research, please ⭐ like the repo and cite it! এই বেঞ্চমার্ক আপনার কাজে লাগলে রেপোতে ❤️ দিন এবং উদ্ধৃত করুন।