CAPEval

CAPEval (Coverage And Precision Evaluation) is a checklist-based caption evaluation benchmark. It decouples caption quality into Coverage (C) and Precision (P) (0–100), and studies how each profile transfers to VLM understanding and T2I generation.

• Code / docs: liuzhipenggg/CAPEval
• Project page: liuzhipenggg.github.io/CAPEval
• Paper: arXiv:2608.02589
• Leaderboard: leaderboard

Dataset contents

Path	Description
image/	300 high-resolution images (up to 8K)
gt_caption.jsonl	Human-written ground-truth captions
checklist.jsonl	Human-verified atomic checklist items (14,965 total)
meta/	Category / label metadata tables

Join key across files: image basename / img_path (e.g. SO001.jpg).

4 super-categories: Scene & Object · People & Activity · Text & Interface · Design & Knowledge.

Metrics (in code)

CAPEval judges each caption against checklist items (yes / no / not_mentioned):

Metric	Definition
C	100 × (yes + no) / total — coverage
P	100 × yes / (yes + no) — precision

Quick start

hf download LiuzhipengUCAS/CAPEval --repo-type dataset --local-dir ./capeval_data

Then point CAPEval env vars at the downloaded paths (see the GitHub README / examples/cluster_run.md).

Citation

@article{liu2026capeval,
  title={CAPEval: A Decoupled Caption Evaluation across Understanding and Generation},
  author={Liu, Zhipeng and Wang, Haochen and Zhang, Zhaoxiang},
  journal={arXiv preprint arXiv:2608.02589},
  year={2026}
}

License

Apache License 2.0 — see the code repository LICENSE.