OmniMedVQA-Diff

Hard-negative Visual Question Answering (VQA) pairs mined from OmniMedVQA, spanning diverse medical imaging modalities.

Each row contains a pair of visually similar images with the same question type but different answers — designed for evaluating and improving model robustness to subtle visual differences.

Processing Pipeline

1. Normalize OmniMedVQA training split (71,334 samples) grouped by imaging modality
2. Embed all images using BiomedCLIP
3. Mine pairs via approximate nearest-neighbor search (FAISS)
4. Filter pairs:
   • Remove pairs with identical images (image_similarity = 1.0)
   • Keep only matching question types
   • Keep only pairs with different answers

Result: 141,982 hard-negative pairs

Schema

Column	Type	Description
pair_id	int	Unique sequential identifier
image_a	Image	Image A (embedded)
image_b	Image	Image B (embedded)
question_a	string	VQA question for image A
question_b	string	VQA question for image B
answer_a	string	Ground-truth answer for image A
answer_b	string	Ground-truth answer for image B
meta	string (JSON)	Nested metadata (see below)

Meta fields

{
  "content_type": "Microscopy Images_train",
  "image_similarity": 0.993,
  "question_similarity": 0.994,
  "image_a": {
    "image_rel": "Images/MHSMA/val/val_145.jpg",
    "source": "Microscopy Images_train",
    "omnimedvqa_index": 50500,
    "sample_id": "omnimedvqa:modality:train:50500",
    "answer_letter": "B",
    "choices": {"A": "...", "B": "..."}
  },
  "image_b": { "..." }
}

Usage

from datasets import load_dataset
import json

ds = load_dataset("JiatanHuang/OmniMedVQA-Diff", split="train")

row = ds[0]
meta = json.loads(row["meta"])
print(row["question_a"], "→", row["answer_a"])
print(row["question_b"], "→", row["answer_b"])
print("Image similarity:", meta["image_similarity"])

Source

Built from OmniMedVQA (modality split). Images are embedded directly in this dataset.