mtybilly/OmniMedVQA-V2

OmniMedVQA — Granular Subsets (v2)

OmniMedVQA is a large-scale medical visual question answering benchmark covering 12 imaging modalities and 5 clinical question types. This v2 release replaces the coarse modality and question_type configs with 13 granular named configs (mod-* and qt-*) whose train/test boundaries follow Med-R1's partitioning. Images are sourced from the canonical foreverbeliever/OmniMedVQA release; restricted-access images (those not distributed in the open-access pool) are present as metadata-only rows with image=None.

---

Abbreviation Key

Config	Full name	Category
mod-ct	CT (Computed Tomography)	Modality
mod-mri	MR (Magnetic Resonance Imaging)	Modality
mod-xray	X-Ray	Modality
mod-fundus	Fundus Photography	Modality
mod-derm	Dermoscopy	Modality
mod-micro	Microscopy Images	Modality
mod-oct	OCT (Optical Coherence Tomography)	Modality
mod-us	Ultrasound	Modality
qt-ai	Anatomy Identification	Question Type
qt-dd	Disease Diagnosis	Question Type
qt-lg	Lesion Grading	Question Type
qt-mr	Modality Recognition	Question Type
qt-oba	Other Biological Attributes	Question Type

---

Per-Subset Row Counts

restricted = rows where image=None (restricted-access source dataset; metadata present).

Config	Train	Test	Restricted (train+test)
mod-ct	12,567	3,241	0
mod-mri	25,507	6,370	0
mod-xray	6,301	1,615	0
mod-fundus	4,300	1,098	0
mod-derm	5,373	1,306	0
mod-micro	4,570	1,110	0
mod-oct	3,798	848	0
mod-us	8,917	2,074	0
qt-ai	13,119	3,329	0
qt-dd	44,329	11,058	0
qt-lg	1,662	436	0
qt-mr	9,173	2,392	0
qt-oba	2,794	704	0
Total	71,333	17,919	0

All Med-R1 rows map exclusively to open-access source datasets; no restricted-access rows (image=None) appear in this release.

Important: mod-* and qt-* subsets are not orthogonal views of the same pool. Med-R1 partitioned them independently with separate train/test boundaries. A given (image, question) pair may appear in a mod-* train split and a qt-* test split (or vice versa). Do not assume the union of all mod-* rows equals the union of all qt-* rows.

---

Schema

Field	Type	Description
image	Image	Full-resolution image bytes. None for restricted-access rows.
problem	string	Med-R1 inline-options format: "<question> A)opt, B)opt, C)opt, D)opt"
solution	string	Med-R1 answer format: "<answer> X </answer>" (spaces inside tags)
question	string	Plain question text with options stripped
choice_a	string	Text of option A
choice_b	string	Text of option B
choice_c	string	Text of option C; "" for binary (Yes/No) rows
choice_d	string	Text of option D; "" if only 2–3 options present (~7.8% of rows)
answer_letter	string	Single char "A" / "B" / "C" / "D"
answer_text	string	Full text of the correct choice
modality	string	Full modality name. Fixed for mod-*; best-effort from foreverbeliever QA for qt-*; "" if not found.
question_type	string	Full question-type name. Fixed for qt-*; best-effort from foreverbeliever QA for mod-*; "" if not found.

---

Example Row

{
  "image": "<PIL.Image 480x353 RGBA>",
  "problem": "What modality is used to capture this image? A)CT, B)X-ray, C)Nuclear medicine scan, D)PET scan",
  "solution": "<answer> A </answer>",
  "question": "What modality is used to capture this image?",
  "choice_a": "CT",
  "choice_b": "X-ray",
  "choice_c": "Nuclear medicine scan",
  "choice_d": "PET scan",
  "answer_letter": "A",
  "answer_text": "CT",
  "modality": "CT(Computed Tomography)",
  "question_type": "Other Biological Attributes"
}

---

Usage

from datasets import load_dataset

# Load CT modality subset — train split
ds = load_dataset("mtybilly/OmniMedVQA-V2", "mod-ct", split="train")
print(ds[0])

# Load Disease Diagnosis subset — test split
ds = load_dataset("mtybilly/OmniMedVQA-V2", "qt-dd", split="test")

# Iterate all 13 configs
configs = [
    "mod-ct", "mod-mri", "mod-xray", "mod-fundus",
    "mod-derm", "mod-micro", "mod-oct", "mod-us",
    "qt-ai", "qt-dd", "qt-lg", "qt-mr", "qt-oba",
]
for cfg in configs:
    ds = load_dataset("mtybilly/OmniMedVQA-V2", cfg)
    print(cfg, ds)

# Filter out restricted-access rows (image=None)
ds_open = ds.filter(lambda x: x["image"] is not None)

For training with max_pixels=401408 (project standard):

from datasets import load_dataset
from PIL import Image as PILImage

def resize_if_needed(example, max_pixels=401408):
    img = example["image"]
    if img is None:
        return example
    w, h = img.size
    if w * h > max_pixels:
        scale = (max_pixels / (w * h)) ** 0.5
        img = img.resize((int(w * scale), int(h * scale)), PILImage.LANCZOS)
    example["image"] = img
    return example

---

Source & Derivation

This dataset is derived from three sources:

1. OmniMedVQA (canonical benchmark):
   Hu, Yutao et al. "OmniMedVQA: A New Large-Scale Comprehensive Evaluation Benchmark for Medical LVLM." NeurIPS 2024. arXiv:2402.09181.
   HF upstream: foreverbeliever/OmniMedVQA

2. Med-R1 (train/test partitioning):
   Lai, Yuxiang et al. "Med-R1: Reinforcement Learning for Generalizable Medical Reasoning in Vision-Language Models." arXiv:2503.13939.
   GitHub: Yuxiang-Lai117/Med-R1 — Splits/{modality,question_type}/{train,test}/*.json

3. This release (mtybilly/OmniMedVQA-V2):
   Restructured from the above two sources. The train/test split boundaries for each of the 13 configs are taken verbatim from Med-R1's JSON files. Image bytes are sourced from foreverbeliever/OmniMedVQA's open-access pool. The problem, solution, modality, and question_type fields use the Med-R1 text and the foreverbeliever QA JSON metadata respectively.

---

License

CC BY-NC 4.0 — Non-commercial use only.

This dataset inherits the license of its upstream sources:

• foreverbeliever/OmniMedVQA is released under CC BY-NC 4.0
• OmniMedVQA is built from multiple publicly available medical datasets, each with their own licenses (see the original OmniMedVQA paper and repository for per-dataset attribution)

Non-commercial use only. Do not use this dataset for commercial purposes.

---

Changelog

Version	Date	Notes
v1	—	Original mtybilly/OmniMedVQA with two configs: modality, question_type
v2	2026-04-16	Replaced modality/question_type configs with 13 granular mod-*/qt-* subsets following Med-R1's partitioning. Added question, choice_a/b/c/d, answer_letter, answer_text fields. Image bytes from foreverbeliever/OmniMedVQA open-access pool.

---

Known Limitations

1. Restricted-access images: All 88,995 rows (71,333 train + 17,662 test across mod-* subsets; 71,077 train + 17,919 test across qt-* subsets) map to open-access source datasets — no image=None rows in this release. (foreverbeliever/OmniMedVQA contains ~35k restricted-access QA records, but Med-R1's partitioning happened to only draw from the open-access pool.)

2. Orthogonal label coverage: The modality field in qt-* subsets and the question_type field in mod-* subsets are populated on a best-effort basis from foreverbeliever's QA JSONs. If a row's image path does not appear in the open-access or restricted-access QA pool, the field is "".

3. Independent partition boundaries: mod-* and qt-* subsets do not cover the same rows. A row from mod-ct/train may appear in qt-mr/test — this is by design, reflecting Med-R1's independent partitioning.

4. Images not resized: Images are stored at original resolution. Apply max_pixels=401408 yourself at training/eval time (project standard).

5. Binary and 3-option rows: ~7.8% of rows have choice_d="" (3-option) and a small fraction have choice_c=choice_d="" (binary Yes/No questions). These are valid rows from the original OmniMedVQA benchmark.