Datasets:

StarPigeon
/

DoveBench

license: gpl-3.0
language:
  - en
  - zh
language_creators:
  - found
  - machine-generated
task_categories:
  - translation
pretty_name: DoveBench
size_categories:
  - 10K<n<100K
tags:
  - long-form
  - subtitles
  - starcraft
  - video-translation
source_datasets:
  - original

DoveBench

A segment-level English→Chinese translation benchmark derived from the ViDove release corpus. DoveBench pairs each Chinese translation segment with an aligned English reference, packaged as a single JSONL so evaluation runs directly without per-file parsing.

Companion dataset for ViDove: A Translation Agent System with Multimodal Context and Memory-Augmented Reasoning (Lu et al., 2025).

Quick stats


Videos	28
Segments	13,704
Total audio	766 min (12.77 h)
Language pair	English → Chinese
Source	sc2 corpus (StarCraft II commentary / interviews)

Note. Original video files are not shipped with this release. Each row carries a youtube_id; access the source video at https://www.youtube.com/watch?v={youtube_id}.

Row schema (`dovebench.jsonl`)

{
  "video_id":   "sc2/27_59",
  "youtube_id": "G3eZhv5qdDo",
  "idx":        0,
  "start":      4.120,
  "end":        7.840,
  "zh_text":    "...",
  "en_text":    "..."
}

Field	Type	Description
`video_id`	`str`	Corpus-qualified ID (`sc2/<key>`).
`youtube_id`	`str`	11-char YouTube video ID — concatenate with `https://www.youtube.com/watch?v=` to view the source.
`idx`	`int`	Zero-based segment index within the video.
`start`, `end`	`float`	Segment boundaries in seconds, sourced from the Chinese subtitle timeline.
`zh_text`	`str`	Chinese translation for this segment.
`en_text`	`str`	Aligned English reference.

How to use

HuggingFace `datasets`

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files="dovebench/dovebench.jsonl",
    split="train",
)
row = next(r for r in ds if r["video_id"] == "sc2/13_07")
print(row)
# → {'video_id': 'sc2/13_07', 'youtube_id': 'n-Dx7MLqFqY', 'idx': 0,
#    'start': 0.0, 'end': 3.55,
#    'zh_text': '欢迎来到这一场的SC2专业比赛',
#    'en_text': 'Welcome back to a professional match of StarCraft 2.'}

Plain Python

import json
with open("dovebench/dovebench.jsonl") as f:
    rows = [json.loads(line) for line in f]

# Group by video
from collections import defaultdict
by_video = defaultdict(list)
for r in rows:
    by_video[r["video_id"]].append(r)

View a source video

row = rows[0]
print(f"https://www.youtube.com/watch?v={row['youtube_id']}&t={int(row['start'])}s")

Citation

@misc{lu2025vidovetranslationagentmultimodal,
      title={ViDove: A Translation Agent System with Multimodal Context and Memory-Augmented Reasoning},
      author={Yichen Lu and Wei Dai and Jiaen Liu and Ching Wing Kwok and Zongheng Wu and Xudong Xiao and Ao Sun and Sheng Fu and Jianyuan Zhan and Yian Wang and Takatomo Saito and Sicheng Lai},
      year={2025},
      eprint={2507.07306},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2507.07306},
}