SCOPE-Bench

Content Depth Matters in Short-Video Recommendation:
Rethinking the Attention Economy

Liwei Deng¹  ·  Jing Jiang¹  ·  Zhiwei Li¹  ·  Allison Clarke²  ·  Yang Wang²  ·  Guodong Long¹

¹ University of Technology Sydney
² Department of Health, Disability and Ageing

🌐 Project Page  ·  💻 GitHub  ·  📄 Paper  ·  🤗 Download

153,561 videos  |  Corrected ASR alignment  |  Content depth evaluation

---

📢 Current release: The preprocessed and alignment-corrected item metadata, items_final_fixed.json, is available now. The final Content Depth Score (CDS) scoring files will be uploaded after the paper is accepted.

🎯 Overview

SCOPE-Bench (Short-video COntent dePth Evaluation Benchmark) studies the depth of information in short-video recommendations, complementing engagement metrics with content depth evaluation.

Perspective	Measure	Purpose
Individual video	CDS	Assess how thoroughly a video develops its information using a seven-level rubric
Recommendation list	LCDS	Evaluate the content depth of a top-k list alongside Recall, NDCG, and Precision

Item-level content depth annotation.

Recommendation evaluation with engagement and content depth metrics.

💾 Released Dataset

File	Records	Format	Size
items_final_fixed.json	153,561	UTF-8 JSON array	≈ 320 MB

The file contains video identifiers, captions, category labels, aligned ASR transcripts, and raw-video references. It is the complete item catalog; the Hugging Face items split does not represent a training or test partition. Video files, interactions, features, and CDS scoring files are not part of this upload.

🛠️ Preprocessing & Alignment Correction

We corrected mismatches between canonical item metadata and raw-video transcripts while preserving canonical identifiers, titles, and categories. Unresolved transcript alignments remain null rather than receiving an unverified match.

For the exact correction code and reconstruction steps, see the GitHub repository, repair scripts, and data repair guide.

🚀 Quick Start

Download and read the released JSON:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="LiweiDeng/SCOPE-Bench",
    filename="items_final_fixed.json",
    repo_type="dataset",
)
with open(path, encoding="utf-8") as f:
    items = json.load(f)

For benchmark setup and evaluation, visit SCOPE-Bench on GitHub.

✒️ Citation

If you use SCOPE-Bench, please cite our paper and acknowledge the original ShortVideo dataset.

@misc{deng2026contentdepthmattersshortvideo,
      title={Content Depth Matters in Short-Video Recommendation: Rethinking the Attention Economy},
      author={Liwei Deng and Jing Jiang and Zhiwei Li and Yang Wang and Guodong Long},
      year={2026},
      eprint={2608.13990},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2608.13990},
}

🙏 Acknowledgements

This release builds on the WWW2025 ShortVideo dataset. We thank its authors for making their data available. The underlying data remain subject to their original licenses and terms.