Datasets:

IndexTeam
/

CASTER-Bench

license: cc-by-nc-4.0
task_categories:
  - video-classification
  - text-classification
language:
  - zh
size_categories:
  - 1K<n<10K
tags:
  - video-quality-assessment
  - user-generated-content
  - social-reasoning
  - community-resonance
  - multimodal
configs:
  - config_name: default
    data_files:
      - path: CASTER_Bench.json
        split: test

CASTER-Bench

CASTER-Bench is a human-annotated multimodal benchmark for Community-Aware Assessment of Social Textual Engagement and Resonance (CASTER) — a task that evaluates whether User-Generated Content (UGC) achieves positive community resonance, going beyond traditional aesthetic-focused Video Quality Assessment (VQA).

This benchmark is introduced in the ACL 2026 paper: "Community-Aware Assessment of Social Textual Engagement and Resonance: A Human-Centric Perspective on User-Generated Content Evaluation".

Motivation

Traditional VQA focuses on pixel-level integrity and low-level visual cues, which fails to capture how quality is actually perceived on UGC platforms. CASTER redefines UGC quality through the lens of social reasoning — assessing whether content elicits genuine community resonance based on its multimodal attributes (title, tags, video content, cover image) rather than visual quality alone.

Dataset Description

CASTER-Bench contains 1,485 long-form UGC videos (average duration: 442 seconds, total: 182.5 hours) sourced from Bilibili, covering 30 major content categories with balanced representation. Unlike existing VQA datasets that rely on short clips (8-10s), CASTER-Bench enables evaluation of narrative coherence, information density, and sustained engagement.

Key Features

Long-form videos: Average 442s duration (vs. 8-20s in prior VQA benchmarks)
Rich multimodal signals: Title, tags, video content, cover image, category metadata
Expert annotations: Labeled by trained content moderators using a human-centered rubric grounded in real community feedback
Diverse categories: 30 top-level categories, 166 sub-categories

Quality Label Distribution

Label	Chinese	Count	Percentage
Excellent	优	158	10.6%
Good	良	253	17.0%
Average	中	573	38.6%
Poor	差	501	33.7%

This distribution mirrors real-world platforms, presenting a realistic challenge for identifying high-quality content amidst massive amounts of average data.

Dataset Structure

CASTER-Bench/
├── README.md
├── CASTER_Bench.json    # Annotations (1,485 entries)
├── cover.zip            # Video cover images
└── video.zip            # Video files

Annotation Format

Each entry in CASTER_Bench.json:

{
  "archive_title": "Video title",
  "tag": "Comma-separated content tags",
  "new_tid_name": "Top-level category",
  "new_sub_tid_name": "Sub-category",
  "video_path": "video/xxxxx.mp4",
  "cover_path": "cover/xxxxx.jpg",
  "human_label": "优/良/中/差"
}

Usage

import json
from huggingface_hub import hf_hub_download

# Download annotations
path = hf_hub_download(repo_id="IndexTeam/CASTER-Bench", filename="CASTER_Bench.json", repo_type="dataset")
with open(path) as f:
    data = json.load(f)

# Download videos and covers
video_zip = hf_hub_download(repo_id="IndexTeam/CASTER-Bench", filename="video.zip", repo_type="dataset")
cover_zip = hf_hub_download(repo_id="IndexTeam/CASTER-Bench", filename="cover.zip", repo_type="dataset")

Comparison with Existing Benchmarks

Dataset	Videos	Avg Duration	Total Duration	Focus	Modality
KoNViD-1k	1,200	8s	2.7h	Aesthetic & Technical	Video Only
LIVE-VQC	585	10s	1.6h	Aesthetic & Technical	Video Only
YouTube-UGC	1,380	20s	7.7h	Aesthetic & Technical	Video Only
FineVD	6,104	8s	13.6h	Aesthetic & Technical	Video Only
CASTER-Bench	1,485	442s	182.5h	Social & Content	Title+Tags+Cover+Video

Citation

@inproceedings{li2026caster,
  title={Community-Aware Assessment of Social Textual Engagement and Resonance: A Human-Centric Perspective on User-Generated Content Evaluation},
  author={Li, Tianjiao and Zhao, Kai and Li, Xiang and Liu, Yang and Sun, Huyang},
  booktitle={Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL)},
  year={2026}
}

License

This dataset is released under CC BY-NC 4.0.