SEM-VQA: Scanning Electron Microscopy Visual Question Answering Corpus

Licensed CC BY 4.0. Fields marked TODO (authors, citation) still need filling in. See Licensing & provenance below — every image here is a figure panel extracted from a published materials-science paper, so attribution is required on redistribution.

Dataset summary

SEM-VQA pairs 16,192 scanning electron microscopy (SEM) images (47,981 individual figure panels) from the materials-science literature with 224,026 generated question–answer pairs spanning four levels of visual reasoning — observation, detection, identification, and interpretation. Images cover ceramics/production materials, Ni-based alloys, and composites. Questions and reference answers were generated by an LLM (gemini-2.5-flash) grounded in per-image captions and metadata, and a random sample was independently audited by human raters across six candidate generator models to validate quality before this generator was selected for the full corpus.

A held-out, stratified 300-image evaluation set curated from this same pool is released separately as SEM-VQA-Bench.

Unique base images	16,192
Total image panels	47,981
Generated QA pairs	224,026
Reasoning levels	4
Candidate generator models audited	6
Human-rated QA pairs	809

Dataset structure

sem-vqa/
├── images/
│   ├── *.jpg                    # 47,981 JPG panels, e.g. ceramics_prod_img10212_A.jpg
│   └── metadata.jsonl            # 224,026 rows: imagefolder-style join of images -> QA pairs
├── details/                     # 47,981 per-panel JSON records (raw subcaption + provenance)
├── stage1/
│   ├── train.json               # 42,989 image–summary pairs (captioning split)
│   ├── validation.json          #  2,345 image–summary pairs
│   └── test.json                #  2,347 image–summary pairs
├── sem_vqa_captions.csv         # 47,981 rows: caption + summary per panel
├── sem_vqa_corpus.jsonl         # 224,026 rows: generated VQA pairs
└── qna_gen_ratings.csv          #    809 rows: human quality audit of the QA generator

sem_vqa_corpus.jsonl — the VQA corpus

One QA pair per line.

Field	Type	Description
question / answer	str	Generated question and reference answer
visual_evidence	str	The visible evidence cited to justify the answer
level	str	Reasoning level: observation, detection, identification, interpretation
feature	str	Free-text microstructural feature the question targets (none if general)
filename / image_name / panel	str	Identify the source image panel and image
prefix	str	Material family prefix (ceramics_prod_, ni_alloy_, composite_, or generic img)
model	str	Generator model (gemini-2.5-flash for the full corpus)
q_index	int	Question index within its image
split	str	train, test, or benchmark — the benchmark rows are the source pool later curated (deduplicated and stratified) into SEM-VQA-Bench, not a benchmark split you should train against
md5	str	MD5 checksum of the source image

Split sizes — 198,764 train / 23,763 test / 1,499 benchmark-pool QA pairs, over 14,381 / 1,752 / 296 unique images respectively.

Reasoning-level distribution (224,026 pairs; 27 malformed multi-label rows <0.01% excluded):

Level	QA pairs	Share
Observation	47,924	21.4%
Detection	97,409	43.5%
Identification	28,744	12.8%
Interpretation	49,922	22.3%

Answer length grows with reasoning complexity for three of the four levels (median words: observation 29, detection 25, identification 19, interpretation 28):

sem_vqa_captions.csv — per-panel captions

Field	Description
image_path	Relative path, e.g. images/ni_alloy_img25315_D.jpg
caption	Short subcaption for this panel
summary	Longer, self-contained description of the panel
visualization_category / visualization_subtype	Almost entirely Microscopy / SEM (one Compositional Map / EDS Map outlier)
image_id	Base image id, e.g. ni_alloy_img25315
panel_suffix	Sub-panel letter (A–R) or single

details/*.json — raw per-panel records

One file per panel ({image_id}_{panel_suffix}.json), holding the original subcaption, summary, and a source_shard field recording provenance from the upstream parquet shard the panel was extracted from. This is the raw material sem_vqa_captions.csv was compiled from.

stage1/{train,validation,test}.json — captioning splits

Simple image → summary pairs (42,989 / 2,345 / 2,347) intended for an image-captioning task rather than VQA. Note: each record also carries a split field internally set to "train" regardless of which file it's in — use the file the record lives in as the authoritative split, not the field.

qna_gen_ratings.csv — generator quality audit

809 QA pairs, each independently rated by a human annotator, used to select gemini-2.5-flash as the corpus generator over five other candidates.

Field	Description
model	Candidate generator model for this sample
grounding / validity / qa_quality	good / average / poor
answerability	yes / partially / no
hallucination	yes / no
blind	Whether the rater saw the model identity while rating
doi	DOI of the source paper the image was drawn from
notes	Free-text rater notes (often empty)

Audit results, overall (n = 809):

Dimension	Good / Yes / No-hallucination	Average / Partially	Poor / No / Hallucinated
Overall QA quality	78.9%	18.3%	2.8%
Visual grounding	91.7%	7.2%	1.1%
Question validity	90.2%	7.9%	1.9%
Answerable from image	94.7%	4.1%	1.2%
Hallucinated content	96.7% (no)	—	3.3% (yes)

QA quality by candidate generator model:

Model	n	Good	Average	Poor
gemini-2.5-flash	156	95.5%	3.8%	0.6%
gemini-3.1-flash-lite	118	85.6%	14.4%	0.0%
gemini-3.5-flash-lite	98	85.7%	13.3%	1.0%
deepseek-v4-flash	137	75.9%	21.2%	2.9%
gpt-5.4-mini	127	67.7%	29.1%	3.1%
mistral-large-3	173	65.9%	26.6%	7.5%

gemini-2.5-flash was selected as the production generator for sem_vqa_corpus.jsonl based on this audit.

Material coverage

Material family is recoverable from the prefix (corpus) or image_id (captions) field. Across the 16,192 unique base images:

Material family	Images	Share
Ceramics / production	8,001	49%
Ni-based alloys	5,597	35%
Composites	691	4%
Other / unlabeled	1,903	12%

How to load

This repo exposes four configs. default decodes actual images inline (best for browsing in the Hub viewer or for training VLMs directly); the others give you the plain, lightweight text tables.

from datasets import load_dataset

# default config: one row per QA pair, `image` decoded as a PIL image
qa_with_images = load_dataset("uhbuvbuvu/sem-vqa")  # config_name="default"

# plain text tables (no image decoding, much lighter to load)
corpus   = load_dataset("uhbuvbuvu/sem-vqa", "raw_corpus")
captions = load_dataset("uhbuvbuvu/sem-vqa", "captions")
ratings  = load_dataset("uhbuvbuvu/sem-vqa", "ratings")

default is backed by images/*.jpg plus images/metadata.jsonl (the imagefolder-with-metadata convention — the same 47,981 image files are shared across all 224,026 rows of metadata, so images are not duplicated on disk; datasets decodes each row's referenced file on the fly). raw_corpus is the same 224,026 rows without image decoding, reading sem_vqa_corpus.jsonl directly — resolve its filename field against images/ yourself if you need pixels:

from PIL import Image
img = Image.open(f"images/{corpus[0]['filename']}")

Data collection process

1. SEM figure panels were extracted from published materials-science papers (tracked per-sample via the doi field in qna_gen_ratings.csv and the source_shard field in details/*.json).
2. Per-panel captions and summaries were produced first (sem_vqa_captions.csv, details/, stage1/).
3. An LLM (gemini-2.5-flash) generated grounded question–answer pairs at four reasoning levels, conditioned on each panel's caption/summary (sem_vqa_corpus.jsonl).
4. Before committing to this generator, six candidate LLMs were compared on a 809-sample human-rated audit (qna_gen_ratings.csv); gemini-2.5-flash scored highest on quality and lowest on hallucination and was selected.
5. A stratified 300-image subset of this corpus was separately curated into the companion SEM-VQA-Bench evaluation set.

Licensing & provenance

This dataset is released under CC BY 4.0. Every image in this dataset is a figure panel extracted from a peer-reviewed materials-science article, identified by DOI (see qna_gen_ratings.csv:doi for a sample; the full per-image DOI mapping should be published alongside this card so downstream users can attribute individual source articles). DOIs observed in this bundle include articles from Journal of Materials Research and Technology, Results in Engineering, and Chinese Journal of Aeronautics — open-access journals published under CC BY 4.0.

Under CC BY 4.0 you are free to share and adapt this dataset, including commercially, provided you give appropriate credit. When redistributing, attribute both this dataset and, where practical, the original source articles by DOI. Captions, summaries, and QA pairs are synthetic (LLM-generated).

Citation

@dataset{TODO_sem_vqa,
  title        = {SEM-VQA: A Visual Question Answering Corpus for Scanning Electron Microscopy},
  author       = {TODO},
  year         = {2026},
  note         = {TODO: venue / arXiv / ICLR submission link}
}