Lruca/VidUEU100k

VidUEU-100K: An Agent-Curated Dataset for Multimodal Understanding, Editing, and Unified Tasks

Introduced in the paper VidUEU-Agent: A Data-Curation Agent for Multimodal Understanding, Editing, and Unified Tasks.

Paper: arXiv:2611.04418 Code: GitHub

VidUEU-100K Dataset Overview

We introduce VidUEU-100K, a large-scale multimodal instruction-tuning dataset constructed by VidUEU-Agent. The agent converts raw videos into traceable metadata and synthesizes instruction samples for multimodal understanding, image editing, and unified interleaved tasks.

VidUEU-100K contains 100K instruction-tuning samples divided into three capability paradigms:

• 40K understanding samples;
• 30K editing samples;
• 30K unified samples.

The release contains 24,025 metadata groups accompanying the 100K instruction samples (the paper reports approximately 20K).

Video sources include OpenVid-1M, Panda-70M, Wikimedia, and YouTube, with FineGym, BDD100K, and CodeVidCL providing additional domain coverage.

Examples of understanding, editing, and unified tasks are shown below.

Download VidUEU-100K Dataset

Install the Hugging Face Hub client and download the repository:

pip install -U huggingface_hub
hf download Lruca/VidUEU100k --repo-type dataset --local-dir VidUEU100k

The paper specifies a planned CC-BY-4.0 release for metadata, annotations, prompts, task JSON files, and redistributable derived data generated by us. Upstream visual assets remain subject to their source licenses, attribution requirements, ShareAlike requirements, and terms of service.

Annotation Format

The original JSONL follows task-specific instruction templates and a shared metadata schema. The HF Parquet configurations provide a reversible compatibility representation; see RELEASE_FORMAT.md.

Instruction subset	Main fields in the paper templates
Understanding	Single-image tasks use image_pos, category, question, answer, and evidence_refs, with options for multiple-choice questions. Multi-image logic uses image_pos_list and ordered turns containing role, text, and evidence_refs.
Editing	from_pos, to_pos, ok, edit_count, edits, instruction, reason, and evidence_refs.
Unified	Task-specific rounds such as Round_Plan, Round_Understand, Round_QA, Round_Overall, Round_Steps, Round_Explain, Round_Gen, and Round_Select, with grounded frame references and evidence_refs.

The minimal metadata schema contains five components:

segment:     video_id, start, end, summary, type
chapters:    chapter_id, time_span, summary, role
shots:       shot_id, chapter_id, time_span, summary, role,
             scene_type, quality_hint, info_hint
keyframes:   frame_id, shot_id, timestamp, caption, quality,
             value_score, tool_semantics, evidence_id
transitions: source, target, atomic_changes, same_shot,
             viewpoint_change, subject_consistent, reliability

Enhance Multimodal Models with VidUEU-100K

Using VidUEU-100K, we LoRA-fine-tune InternVL3-8B on the understanding subset and Bagel-7B-MoT separately on each of the three capability subsets. The tables below report the corresponding benchmark results from the paper.

Results on Understanding Tasks

Model	Size	MMT-Bench	MME	LLaVA-Bench	MMDU
InternVL3	8B	0.546	2308.0	69.0	53.1
Ours (InternVL3)	8B	0.596	2369.3	70.8	61.1
Bagel	14B	0.645	1573.6	47.5	28.5
Ours (Bagel)	14B	0.651	1900.2	57.9	33.2

Results on Editing Tasks (GEdit-Bench-En)

Model	Semantic Consistency	Perceptual Quality	Overall
Bagel	7.358	6.831	6.527
Ours (Bagel)	7.565	7.328	6.951

Results on Unified Tasks (WEAVEbench)

Method	Science	Creation	Logic	Game	Overall
Bagel	0.378	0.475	0.406	0.365	0.446
Bagel-Zebra	0.399	0.456	0.393	0.396	0.449
Ours (Bagel)	0.426	0.501	0.480	0.419	0.501

Citation

If you find VidUEU-100K or VidUEU-Agent useful, please cite our paper:

@inproceedings{ru2026vidueuagent,
  title     = {{VidUEU-Agent}: A Data-Curation Agent for Multimodal Understanding, Editing, and Unified Tasks},
  author    = {Ru, Tengju and Lian, Weitong and Tang, Zecong and Meng, Lingyi and Li, Haoran and Cui, Zhejun and Zhu, Yichen and Cao, Hangshuo and Kang, Qi and Liu, Yechi and Wang, Kaixuan and Yuan, Yu-Jie and Wang, Chunwei and Zhang, Yu and Dai, Bo},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}