Wiki-CoE

Wiki-CoE is a multimodal question-answering dataset for evaluating whether a model can answer questions from visual evidence and produce an explicit chain of evidence. Each example pairs a natural-language question with one or more Wikipedia page screenshots and asks the model to return both the answer and the supporting evidence.

The dataset is intended for research on visual reasoning, evidence localization, and chain-of-evidence style multimodal QA.

Dataset contents

The main release is under release_strict_chainsplit_20260428_1338/:

File	Description
train.json	training split, 35,210 examples
test.json	test split, 35,208 examples
val_1000_seed42.json	small validation subset, 1,000 examples
split_report.json	high-level split statistics
val_1000_seed42_manifest.json	validation subset indices

Each example contains:

Field	Description
question	input question
answer	short answer
type	reasoning type
num_hops	number of evidence hops
evidence_chain	ordered evidence screenshots, bounding boxes, and sub-questions
evidences	supporting evidence metadata

Screenshot archive

The screenshot images are stored as a split archive under release_assets/.

To reconstruct the image directory:

cat release_assets/screenshots_trimmed.tar.zst.part-* > screenshots_trimmed.tar.zst
tar --use-compress-program=unzstd -xf screenshots_trimmed.tar.zst

This creates:

screenshots_trimmed/

The archive contains 76,000 PNG screenshots. SHA256 checksums for the archive parts are provided in release_assets/screenshots_trimmed.tar.zst.sha256.

Related resources

• Homepage: https://lpy.pxsec.cn
• Paper: https://arxiv.org/abs/2605.01284
• Code: https://github.com/PeiYangLiu/CoE
• Wiki-CoE model checkpoint: https://huggingface.co/PeiyangLiu/CoE-Wiki-CoE-8B
• SlideVQA model checkpoint: https://huggingface.co/PeiyangLiu/CoE-SlideVQA-8B