Datasets:

tsinghua-ee
/

DVD-Bench

license: cc-by-nc-4.0
task_categories:
  - video-text-to-text
language:
  - en
pretty_name: DVD-Bench
size_categories:
  - 1K<n<10K
tags:
  - video
  - dialogue
  - speaker
  - benchmark
configs:
  - config_name: en
    data_files:
      - split: test
        path: data/en/test.parquet

DVD-Bench

A benchmark for Dialogue-centric Video Description, evaluating "When, Who, and What is Said" in dialogue-centric videos.

Repository layout

DVD-Bench/
├── data/
│   └── en/
│       └── test.parquet       # English test split annotations
└── videos/
    └── en/
        └── *.mp4              # English test split videos

Annotation schema

Each row in data/en/test.parquet contains:

field	type	description
`video`	`string`	Video filename, e.g. `e-mNCJPxQvQ.mp4`. The actual file is at `videos/en/<video>` in this repo.
`character`	`list<string>`	List of character descriptions appearing in the video.
`dialogue`	`list<struct{speaker: string, content: string, time: list<string>}>`	Ordered list of utterances; `time = [start, end]` in `MM:SS`.

Quickstart

from datasets import load_dataset

ds = load_dataset("tsinghua-ee/DVD-Bench", name="en", split="test")
print(ds[0])

# To get the actual video file, download it from videos/en/<video>:
from huggingface_hub import hf_hub_download
video_path = hf_hub_download(
    repo_id="tsinghua-ee/DVD-Bench",
    repo_type="dataset",
    filename=f"videos/en/{ds[0]['video']}",
)

Evaluation Results

Model	DVD-Bench (En) Acc% ↑	DVD-Bench (En) WER% ↓	DVD-Bench (En) IoU% ↑	DVD-Bench (Zh) Acc% ↑	DVD-Bench (Zh) CER% ↓	DVD-Bench (Zh) IoU% ↑
ARC-Qwen-Video-Narrator (7B)	66.4	65.0	23.0	63.2	53.6	10.1
Qwen2.5-Omni (7B)	62.7	83.6	-	55.7	69.4	-
video-SALMONN 2+ (7B)	66.6	94.0	-	59.9	-	-
AVoCaDO (7B)	72.9	17.9	-	69.3	-	-
Qwen3-Omni-Instruct (30B-A3B)	67.8	91.3	-	63.5	60.6	-
Ours-SFT Model (8B)	71.2	29.8	31.8	69.6	30.3	24.7
D-ORCA (8B)	81.1	16.6	57.1	78.0	17.5	37.8

📅 Roadmap

Release DVD-Bench (en) evaluation data.
Release DVD-Bench (zh) evaluation data.
Release DVD-Train dataset annotations.

Reference

If you find DVD-Bench useful for your research, please cite our paper:

@article{tang2026dorca,
  title={{D-ORCA: Dialogue-Centric Optimization for Robust Audio-Visual Captioning}}, 
  author={Changli Tang and Tianyi Wang and Fengyun Rao and Jing LYU and Chao Zhang},
  journal={arXiv preprint arXiv:2602.07960},
  year={2026}
}