elichen-skymizer/vlm-benchmark-questions

VLM Benchmark Questions

1,800 questions and 2,105 image entries across six subsets, each containing the first 300 rows of the corresponding *-subsample-500 subset in elichen-skymizer/vlm-prepared-dataset, pinned at 6cb6a4d65fcb2c8d788f68aaa0aa92c5ceca408b.

Selection is exactly source.select(range(300)), using the original row order. No reshuffling or resampling is performed. Each subset includes the first 100 pilot questions and the next 200 questions from the collection portion. Original question text, image bytes and ordering, reference answers, identifiers, and source metadata are unchanged. Parquet files are rewritten to contain only those rows.

Subset	Questions	Image entries	Original benchmark
blink-val-subsample-300	300	605	BLINK-Benchmark/BLINK
mmmu-pro-vision-subsample-300	300	300	MMMU/MMMU_Pro
mmstar-subsample-300	300	300	Lin-Chen/MMStar
ocrbench-v1-subsample-300	300	300	echo840/OCRBench
ocrbench-v2-subsample-300	300	300	lmms-lab/OCRBench-v2
mathvision-test-subsample-300	300	300	MathLLMs/MathVision

The split name remains train as in the prepared source; this is a packaging convention for benchmark questions.

Usage

from datasets import load_dataset

ds = load_dataset(
    "elichen-skymizer/vlm-benchmark-questions",
    "blink-val-subsample-300",
    split="train",
)
assert len(ds) == 300
example = ds[0]
print(example["item_id"], example["question"], example["ref_answer"])
example["images"][0].show()

The columns are source, item_id, origin_id, sample_seed, question, images, num_images, ref_answer, and category. images is an ordered list of embedded images. ref_answer is the original benchmark reference answer. In MMMU-Pro vision, the question and choices are inside the image, so its question text is empty by design.

Provenance

The original full validation traced 35,320 rows across nine elichen-skymizer/*-collect-400-extended repositories and 90 configurations, checking question metadata and every image byte hash. This selection retains 17,475 collection-row mappings and 20,479 collection image entries, restricted to source row indices below 300. All 1,800 output questions were checked against the source slice and the previously verified question/image hashes, with zero mismatches.

• provenance/source.json: pinned source revision, original source checksums, output checksums, and row selection.
• provenance/questions.jsonl: one record per retained question with source indices and question/image SHA256 values.
• provenance/collection_rows.csv: collected model rows that refer to retained questions. source_config refers to the original subsample-500; question_config refers to this dataset's subsample-300.
• provenance/collections.json: unchanged collection revisions/file hashes and selected row/image counts.
• provenance/validation.json: selection validation and the pinned prior full-validation revision.

All provenance row indices are zero based. Join by configuration and item_id; collection row numbers can shift after model-specific exclusions. The current dataset has six subsample-300 configurations; the previous 500-question files are retained only in repository history.

Attribution

The immediate source is vlm-prepared-dataset; the benchmark links above provide upstream attribution. Preparation may have changed upstream question formatting. All rights to images, questions, and reference answers remain with their original owners, and this compilation does not grant additional rights or relicense their content.

The prepared source documents an image-diversity update and a sampling replacement ledger. The pinned source files and explicit row selection identify the exact questions; sample_seed alone does not describe subsequent replacements.