Accio-Lab/VDR-EE-Benchmark

VDR-EE: Entity- and Event-Centric Video Deep Research

VDR-EE is the evaluation benchmark introduced in AdaVDR: Adaptive Tool Use and Reflection for Video Deep Research. Its 250 manually verified questions connect entities or events in videos to facts from external sources, testing joint video understanding and knowledge retrieval.

Paper (PDF) · GitHub · Project page

This repository provides one test split in Parquet format, containing questions, reference answers, category labels, video source links, and external evidence URLs. Video files are not included. The original question order is preserved.

What does VDR-EE contain?

The benchmark covers two complementary forms of video deep research:

• Entity-centric questions (153): identify a target entity or connect multiple entities appearing in a video, then retrieve the external facts needed to answer the question. The categories are Single Entity and Multiple Entity.
• Event-centric questions (97): recognize and reason about an event shown in a video, then connect it to external knowledge. The categories are Short Event, Medium Event, and Long Event, based on the target event's duration.

Each released record includes a question, reference answer, stable identifier, domain and question-type labels, the source video URL and identifier, an optional video title, and a list of external evidence URLs. The seven domains are Culture, Entertainment, Industrial, News, Scene understanding, Science, and Sports.

Video access and copyright

Due to copyright considerations, this repository distributes annotations and source links only, rather than the original video files. Video rights remain with their respective copyright holders.

To obtain the videos, download them from the original sources listed in the youtube_url column, subject to the source platform's terms and applicable permissions. For assistance with video access, please contact zhangxt119@gmail.com.

Providing a source link does not grant a license to redistribute the associated video. Source availability may change over time.

Dataset statistics

Original figure from the AdaVDR repository. The figure's event-subcategory counts differ from the labels in this Parquet release; use the Parquet labels when computing category-specific metrics.

Representative questions

Illustrative benchmark examples from the AdaVDR project. The figure shows the five question categories and how video observations are combined with external facts.

Columns

Column	Type	Description
uid	string	Stable UUIDv5 derived from the original question ID.
question	string	Question text, preserved from the source.
answer	string	Reference answer, preserved from the source.
domain	string	Culture, Entertainment, Industrial, News, Scene understanding, Science, or Sports.
type	string	Event or Entity.
finegrained_type	string	Long Event, Medium Event, Short Event, Single Entity, or Multiple Entity.
youtube_url	nullable string	Canonical source video URL; null if the source cannot be established.
video_id	string	Original video identifier, preserved as supplied; some identifiers are cache hashes.
video_title	nullable string	Title supplied by the source; missing titles remain null.
evidence_sources	list of strings	External evidence URLs, preserved as a native list.

The domain and both type fields remain separate columns for filtering. Local video paths and internal review records are excluded from this export.

Load

from datasets import load_dataset

dataset = load_dataset("Accio-Lab/VDR-EE-Benchmark", split="test")

example = dataset[0]
print(example["question"])
print(example["youtube_url"])

For a downloaded local copy:

dataset = load_dataset("parquet", data_files={"test": "data/*.parquet"}, split="test")

Evaluation

Use question as the query, retrieve the associated video using youtube_url, and compare the resulting answer with answer. Report overall performance and breakdowns by type, finegrained_type, and domain as appropriate. Reference answers and evidence_sources support evaluation; they should not be supplied to an agent as part of its question input.

Evaluation code and running instructions: Accio-Lab/AdaVDR on GitHub. The project's evaluation code is released through GitHub separately from this Hugging Face dataset. Follow that repository for the evaluation implementation, environment setup, and execution instructions as they become available.

As of October 20, 2026, the GitHub README still lists the benchmark and evaluation code release as pending. This dataset card therefore links to the project without assuming that runnable evaluation scripts are already public.

Benchmark results

The following accuracy scores (%) are reported by the AdaVDR project, using GPT-5.4 as the semantic answer judge in the agentic setting. They are published project results, not a new evaluation run for this dataset upload.

Model	VDR-EE Entity	VDR-EE Event	VDR-EE Overall	VideoDR Overall
Qwen3.5-9B	42.48	36.08	40.00	37.00
AdaVDR-9B-SFT	47.06	49.48	48.00	52.00
AdaVDR-9B	50.33	62.89	55.20	66.00

Compared with the agentic Qwen3.5-9B baseline, AdaVDR-9B improves overall accuracy by 15.20 percentage points on VDR-EE and 29.00 points on VideoDR.

Complete reported results

Original results table from the AdaVDR repository, including other baselines and question-subtype breakdowns. Event-subtype results retain the published grouping; do not recompute their aggregates using this release's different event-subcategory counts.

Citation

If you use VDR-EE in your research, please cite the AdaVDR paper:

@article{zhang2026adavdr,
  title={AdaVDR: Adaptive Tool Use and Reflection for Video Deep Research},
  author={Zhang, Xintong and Fan, Xiaomeng and Yan, Shilin and He, Ekko and Liu, Zicheng and Zou, Zijian and Zhang, Guannan and Wu, Yuwei and Gao, Zhi and Xue, Hongwei},
  journal={arXiv preprint arXiv:2608.25559},
  year={2026}
}

Contact

Questions about the benchmark or video access: zhangxt119@gmail.com.