SEM-VQA-Bench: A Stratified Evaluation Benchmark for SEM Visual Question Answering

Licensed CC BY 4.0 — see Licensing & provenance. This benchmark is a curated subset of SEM-VQA; read that card for full corpus-level context.

Dataset summary

SEM-VQA-Bench is a compact, stratified evaluation set of 1,497 question–answer pairs over 300 scanning electron microscopy (SEM) images, drawn from the larger SEM-VQA corpus. Images were sampled to be representative rather than convenient: they were embedded, clustered into 5 groups, and sampled evenly across core / mid / peripheral distance-to-centroid bands within each cluster — so the benchmark spans both prototypical and atypical images in roughly equal measure, avoiding an evaluation set biased toward easy, cluster-center examples.

The table is one row per image (300 rows), with every question/answer for that image packed into parallel list columns — so each image is stored and rendered in the Hub viewer exactly once.

Images (rows)	300
QA pairs	1,497
Embedding clusters	5
Centrality bands	3 (core / mid / peripheral)
Reasoning levels	4

Dataset structure

A single test split (data/test-*.parquet), one row per image.

Main table — one row per image

Field	Type	Description
id	str	Panel filename, e.g. ceramics_prod_img12266_D.jpg
image	Image	The SEM panel, decoded inline
image_name	str	Base image id (no panel suffix)
questions	list[str]	All benchmark questions for this image
answers	list[str]	Reference answers, parallel to questions
evidence	list[str]	Visual evidence supporting each answer
levels	list[str]	Reasoning level per question: observation / detection / identification / interpretation
features	list[str]	Microstructural feature targeted per question
cluster_id	int	Embedding cluster (0–4) this image was assigned to
centrality	str	core, mid, or peripheral — distance-to-centroid band within its cluster
distance_to_centroid	float	Raw embedding distance to the cluster centroid (mean 0.576, range 0.385–0.838)
n_copies_in_dataset	int	How many times this exact image appears across the full SEM-VQA corpus (1–4)
source_split	str	train or test — the split this image's QA pairs originally belonged to in the source SEM-VQA corpus, kept for traceability. This is not a train/test split of the benchmark itself: SEM-VQA-Bench is a single evaluation set and should not be further split for training.
md5	str	MD5 checksum of the source image

Stratified design balance

Reasoning level (1,497 QA pairs):

Level	QA pairs
Observation	297
Detection	669
Identification	200
Interpretation	331

Cluster centrality (300 images, near-even by design):

Centrality	Images
Core	497
Mid	500
Peripheral	500

Clusters are close to evenly sized too (300 / 300 / 300 / 299 / 298 images across the 5 clusters). Most images carry more than one QA pair (mean 1.65, median 2, max 4 per image — see questions/answers list length, or n_copies_in_dataset).

How to load

from datasets import load_dataset

ds = load_dataset("uhbuvbuvu/sem-vqa-bench", split="test")

row = ds[0]
row["image"]         # PIL.Image
row["questions"][0]  # first question for this image
row["answers"][0]    # its reference answer

To flatten back into one row per QA pair:

flat_rows = [
    {"id": r["id"], "image": r["image"], "question": q, "answer": a, "level": l}
    for r in ds
    for q, a, l in zip(r["questions"], r["answers"], r["levels"])
]

Intended use

This benchmark is meant for held-out evaluation only (e.g. zero-shot / few-shot VQA benchmarking of vision-language models on scientific microscopy imagery). Because its images are drawn from the same pool as the SEM-VQA training corpus, do not train on SEM-VQA using images that also appear here (check md5 / id against your training split) if you intend to report clean held-out numbers.

Data collection process

SEM-VQA-Bench was curated from the SEM-VQA corpus by:

1. Embedding candidate images and clustering them into 5 groups.
2. Computing each image's distance to its cluster centroid and binning into core / mid / peripheral bands.
3. Sampling ~100 images per cluster, balanced across centrality bands, for 300 images total.
4. Carrying over all QA pairs generated for those 300 images (1,497 total), preserving level, feature, and provenance metadata from the parent corpus.

See the SEM-VQA dataset card for how the underlying QA pairs were generated and quality-audited.

Licensing & provenance

This dataset is released under CC BY 4.0, inherited from the parent SEM-VQA dataset — see that card's Licensing & provenance section for the source-article details. Every image referenced here is a figure panel extracted from a peer-reviewed materials-science article; attribute both this dataset and, where practical, the original source articles by DOI when redistributing.