KorMedMCQA-V: A Multimodal Benchmark for Evaluating Vision-Language Models on the Korean Medical Licensing Examination

KorMedMCQA-V is a multimodal multiple-choice question answering benchmark for evaluating vision-language models on the Korean Medical Licensing Examination. The dataset consists of 1,534 questions with 2,043 associated medical images from Korean Medical Licensing Examinations (2012-2023).

Dataset Summary

• Total Questions: 1,534
• Total Images: 2,043 (avg 1.33 images/question)
• Splits: test (2022-2023, 304 questions), test_full (2012-2023, 1,534 questions)
• Format: Parquet with base64-encoded images
• Image Modalities (9 categories): X-ray (586), Other (554), CT (336), ECG (164), Ultrasound (138), Endoscopy (122), NST (54), PBS (49), MRI (40)

Data Format

Each sample contains:

Field	Type	Description
subject	string	Subject type (always "doctor")
year	int64	Year of examination
period	int64	Period of examination
q_number	int64	Question number
question	string	Question text
A, B, C, D, E	string	Answer choices
answer	string	Correct answer (A-E)
images	string	JSON string of base64-encoded image objects

Image Object Structure

images field is a JSON string containing an array of image objects with base64-encoded images:

[
  {
    "pic_num": "1",
    "modality": "XRAY",
    "image_base64": "data:image/png;base64,<base64>"
  }
]

The image_base64 field contains a full data URL.

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("seongsubae/KorMedMCQA-V", name="doctor", split="test_full")

for sample in dataset:
    print(f"Question: {sample['question']}")
    print(f"Answer: {sample['answer']}")

Viewing Images

import json
import base64
import io
from PIL import Image

sample = dataset[0]
images = json.loads(sample["images"])

for img in images:
    data_url = img["image_base64"]
    header, b64_str = data_url.split("base64,", 1)
    img_bytes = base64.b64decode(b64_str)
    pil_image = Image.open(io.BytesIO(img_bytes))
    pil_image.show()
    print(f"Modality: {img['modality']}, Size: {pil_image.size}")

Combining with KorMedMCQA

To evaluate on both text-only and image-dependent questions, combine the test split with sean0042/KorMedMCQA:

• KorMedMCQA (text-only) contains 2022-2024 data; filter to 2022-2023 for alignment
• KorMedMCQA-V (multimodal) contains 2022-2023 data
• Remove duplicate UID doctor-2022-2-64 to avoid double-counting

from datasets import load_dataset

# Load both datasets (test split = 2022-2023)
kormedmcqa = load_dataset("sean0042/KorMedMCQA", name="doctor", split="test")
kormedmcqa_v = load_dataset("seongsubae/KorMedMCQA-V", name="doctor", split="test")

# Filter KorMedMCQA for 2022-2023 and remove duplicate UIDs
allowed_years = [2022, 2023]
excluded_uids = ["doctor-2022-2-64"]

kormedmcqa_filtered = [
    s for s in kormedmcqa
    if s["year"] in allowed_years
    and f"{s['subject']}-{s['year']}-{s['period']}-{s['q_number']}" not in excluded_uids
]

print(f"Text-only: {len(kormedmcqa_filtered)}, Multimodal: {len(kormedmcqa_v)}")

For evaluation code, see the GitHub repository.

License

This dataset is licensed under CC BY-NC-SA 4.0.

Citation

@dataset{kormedmcqa-v,
  title        = {KorMedMCQA-V: A Multimodal Benchmark for Evaluating Vision-Language Models on the Korean Medical Licensing Examination},
  author       = {Byungjin Choi and Seongsu Bae and Sunjun Kweon and Edward Choi},
  year         = {2025},
  publisher    = {HuggingFace},
  version      = {1.0},
 }

Contact

For questions or issues, please contact Byungjin Choi (choi328328@ajou.ac.kr) or Seongsu Bae (seongsu@kaist.ac.kr).