license: cc-by-nc-4.0
task_categories:

• visual-question-answering
• image-text-to-text
  language:
• en
  tags:
• medical
• kvasir
  dataset_info:
  features:
  • name: image
    dtype: string
  • name: complexity
    dtype: int64
  • name: question
    dtype: string
  • name: answer
    dtype: string
  • name: original
    dtype: string
  • name: question_class
    list: string
  • name: img_id
    dtype: string
    splits:
  • name: train
    num_bytes: 71692651
    num_examples: 143594
  • name: test
    num_bytes: 7944880
    num_examples: 15955
    download_size: 16580777
    dataset_size: 79637531
    configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*
  • split: test
    path: data/test-*

Kvasir-VQA-x1

A Multimodal Dataset for Medical Reasoning and Robust MedVQA in Gastrointestinal Endoscopy

Kvasir-VQA-x1 on GitHub |
Original Image from Kvasir-VQA(Simula Datasets) |
Paper

🔗 MediaEval Medico 2025 Challenge uses this dataset. We encourage you to check out and participate!

Overview

Kvasir-VQA-x1 is a large-scale dataset designed to benchmark medical visual question answering (MedVQA) in gastrointestinal (GI) endoscopy. It introduces 159,549 new QA pairs stratified by clinical complexity, along with support for visual robustness testing via augmentations.

Features

Each dataset entry includes:

• img_id: Unique reference to an image from Kvasir-VQA
• complexity: Question complexity level (1–3)
• question: Complex, natural language question
• answer: Clinically validated answer
• original: List of atomic QA pairs merged into the complex question
• question_class: Associated clinical category labels

Splits

• train: Training samples only
• test: Held-out samples for final evaluation

To ensure generalization, no image or QA from the test set appears in the training set.

Downloading Images and Preparing JSONLs

from datasets import load_dataset
from pathlib import Path
from tqdm import tqdm
import pandas as pd
import json
import os

Folder to save images and JSONLs
d_path = "./Kvasir-VQA-x1/"
img_dir = Path(os.path.abspath(os.path.join(d_path, "images")))
img_dir.mkdir(exist_ok=True, parents=True)
Save images once from SimulaMet-HOST/Kvasir-VQA
ds_host = load_dataset("SimulaMet-HOST/Kvasir-VQA", split="raw")
seen = set()
for row in tqdm(ds_host, desc="Saving images"):
    if row["img_id"] not in seen:
        row["image"].save(img_dir / f"{row['img_id']}.jpg")
        seen.add(row["img_id"])
Save VLM-ready JSONLs for SimulaMet/Kvasir-VQA-x1
for split in ["train", "test"]:
    with open(f"{d_path}/Kvasir-VQA-x1-{split}.jsonl", "w", encoding="utf-8") as f:
        for r in load_dataset("SimulaMet/Kvasir-VQA-x1", split=split):
            f.write(json.dumps({
                "messages": [
                    {"role": "user", "content": f"<image>{r['question']}"},
                    {"role": "assistant", "content": r["answer"]}
                ],
                "images": [str(img_dir / f"{r['img_id']}.jpg")]
            }, ensure_ascii=False) + "\n")

🔗 Instructions and code for creating VLM-ready JSONLs that incorporates the augmented images can be found here.

Use Cases

• Multimodal clinical reasoning
• Robustness evaluation under visual perturbations
• Fine-tuning and benchmarking of VLMs (Vision-Language Models)

License

Released under CC BY-NC 4.0 – for academic and non-commercial use. Please cite appropriately.

Citation & Paper

Please cite the associated dataset paper if you use Kvasir-VQA-x1 in your work:

@article{Gautam2025Jun,
    author = {Gautam, Sushant and Riegler, Michael A. and Halvorsen, P{\aa}l},
    title = {{Kvasir-VQA-x1: A Multimodal Dataset for Medical Reasoning and Robust MedVQA in Gastrointestinal Endoscopy}},
    journal = {arXiv},
    year = {2025},
    month = jun,
    eprint = {2506.09958},
    doi = {10.48550/arXiv.2506.09958}
}

Related Repositories

You can find full training scripts, augmentation tools, and baseline models at:
👉 GitHub: simula/Kvasir-VQA-x1