CRPE — re-hosted for lmms-eval

A bundled-image re-host of OpenGVLab/CRPE prepared for upstream-friendly use with EvolvingLMMs-Lab/lmms-eval.

What changed vs. OpenGVLab/CRPE

The original repo ships the crpe_relation.jsonl annotation file alongside 544 abnormal-scene images in abnormal_images/. The remaining 5,400 question records reference COCO val2017 images by relative path — but those JPEGs are not part of the HF repo, so out-of-the-box load_dataset(...) cannot resolve them.

This re-host:

1. Bundles all 1,081 unique referenced images (537 from COCO val2017 + 544 from abnormal_images/) into the parquet itself, encoded as JPEG bytes under an Image() feature.
2. Preserves every question record, choice, target, and category from the original crpe_relation.jsonl.

Result: a self-contained parquet that loads end-to-end via standard load_dataset — no extra COCO download needed.

Schema

Field	Type	Notes
question_id	string	Original question_id from the source jsonl
image	Image()	JPEG-encoded image
text	string	Question + A./B./C./D. options + final instruction
correct_option	string	A / B / C / D
category	string	One of subject, predicate, object

Stats

• 7,576 test items, single test split.
• 1,081 unique images: 537 from COCO val2017, 544 from abnormal_images/.
• Category distribution: subject 3,176, predicate 2,176, object 2,224.
• Parquet payload: ~1.0 GB across 4 shards.

License

This re-host inherits its licensing from the original sources:

• CRPE annotations — © OpenGVLab, released under the original repository's license (see OpenGVLab/CRPE).
• COCO val2017 images — © COCO Consortium, CC BY 4.0.
• Abnormal scene images — bundled from OpenGVLab/CRPE.

Citation

@article{wang2024crpe,
  title={The All-Seeing Project V2: Towards General Relation Comprehension of the Open World},
  author={Wang, Weiyun and others},
  journal={arXiv preprint arXiv:2402.19474},
  year={2024}
}