xiongyq/SCOPED-Eval

SCOPED Evaluation Collection

This dataset is part of the paper Skill-Decoupled On-Policy Expert Distillation for Unified Vision-Language Models. The official code is available at https://github.com/NEUIR/SCOPED.

A private preparation collection of the eight locally used SCOPED evaluation datasets. This repackages original annotations and image bytes; it does not contain model predictions or training augmentations.

Configuration	Source split	Rows	Missing local images
CCOCRV2	test	1790	0
ChartQA	test	2500	0
CountBenchQA	test	491	0
MMSI-Bench	test	1000	0
MUIRBENCH	test	2600	0
pixmo-count	test	540	13
VisualPuzzles	train	1168	0
We-Math	testmini	1740	0

Load

from datasets import load_dataset
data = load_dataset("xiongyq/SCOPED-Eval", "ChartQA", split="test")

Authenticate with an authorized HF account for this private repository. The images column is a list of HF Image features with embedded bytes, including multi-image examples.

Fields and evaluation scope

question, answers, options, task and subset expose annotations. metadata contains the original non-image fields as JSON, including original IDs, option text, grouping keys, source files and split names. Answer strings are retained without rewriting. Image order and original bytes are preserved. Some benchmarks require composing question and options before inference; answers/metadata must not be appended to model prompts.

VisualPuzzles labels its benchmark split train; it is retained under that name here. We-Math uses testmini, also retained. No split is renamed to suggest a new held-out set.

CCOCRV2 includes the recognition track only. The existing SCOPED evaluation script tested recognition only (1,790 examples); packaging additional tracks does not claim results on them. Select rows with task == 'recognition' for that protocol.

PixMo-Count has 540 annotations but 13 images are absent from the supplied local image cache. All rows are retained; those 13 have image_available=False, an empty image list and skip_reason='missing_local_image'. This matches the existing evaluator's missing-image policy. No unavailable image is fabricated or substituted.

Sources, attribution and licensing

Original dataset cards are preserved under sources/. Their licenses apply separately; the SCOPED code's Apache-2.0 license does not license these images or annotations. The original cards record mixed terms, including GPL-3.0, MIT, CC-BY-4.0, ODC-BY and CC-BY-NC-4.0. No blanket license is asserted for this collection. Review original image rights and attribution before a public release.

manifest.json records source file hashes, exported shard hashes and exact row counts. sources/We-Math/knowledge_structure_nodes.json retains the benchmark's auxiliary knowledge metadata.