Crayotter: Learning Long-Horizon Video Editing Agents via Group-Relative Preference Backpropagation
Paper • 2608.02694 • Published
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Please describe your research use and acknowledge the terms in LICENSE.md.
Log in or Sign Up to review the conditions and access this dataset content.
LongEditBench
LongEditBench evaluates long-horizon video editing and contextual revision. An agent uses a bound long source video to produce an approximately 30-second, source-grounded edit. Follow-up requests test whether it can revise its own artifact while preserving applicable earlier requirements.
The benchmark adapts source-grounded interactions from LongShOTBench into executable editing briefs and is compatible with Crayotter.
| Statistic | Value |
|---|---|
| Editing requests | 337 |
| Source videos | 21 |
| Native interactions / initial requests | 253 |
| Initial-follow-up pairs | 84 |
| Total source duration | 12.91 hours |
| Source duration range | 30.37-53.69 minutes |
| Target artifact duration | Approximately 30 seconds |
Data and access
This release provides four Parquet configurations, all under the test split:
tasks: editing briefs, allowed tools, source paths, and links between initial and follow-up requests.references: scorer-only reference answers and task-specific evidence criteria.revision_pairs: the 84 matched initial-follow-up pairs.sources: video identifiers, source URLs, durations, and media metadata.
Task fields include case_id, source_video_id, native_row_index, question_turn_index, task_category, request_type, previous_case_id, user_request, target_duration_seconds, allowed_tools, and media_path. Case names use source, native row, and turn, making the corresponding interaction easy to locate.
This full release includes the 21 source-video binaries under media/ (approximately 20.58 GB / 19.17 GiB in total). The original package files and runtime helpers are retained under package/ for reproducibility. See media/README.md for the mapping from source_video_id to the local video path.
Access to the dataset is intended to use manual approval on the dataset page. After approval, authenticate with hf auth login and load:
from datasets import load_dataset
repo_id = "witness0411/LongEditBench"
tasks = load_dataset(repo_id, "tasks", split="test", token=True)
pairs = load_dataset(repo_id, "revision_pairs", split="test", token=True)
print(tasks[0]["user_request"])
The corresponding source video is available locally at media/<source_video_id>.<ext>. The sources configuration records the relative path and media metadata. The release contains the held-out evaluation package only; it does not add a training split.
For a downloaded repository, no Hub connection is needed:
tasks = load_dataset(
"parquet", data_files={"test": "data/tasks.parquet"}, split="test"
)
Evaluation
Supply each agent with its brief and bound source video. Keep references outside the agent workspace. For contextual revision, execute the initial request, then apply its follow-up to the same agent's artifact and retained project state.
Artifact quality is rated from 1 to 5 on theme fit, content richness, narrative coherence, editing fluency, and visual quality. Report completion and completion-adjusted quality over the full evaluated set; missing or invalid exports receive zero. Contextual revision additionally measures new-request satisfaction, preservation of earlier requirements, and source grounding. Use source-video clusters for statistical comparisons, keeping both turns of a pair together.
These cases are held-out evaluation data. The release contains no training split. Report the evaluator model, prompt, decoding settings, and dataset version when comparing systems. Coverage is limited to the 21 source videos represented here.
License and provenance
Use is subject to LICENSE.md and the original annotation and source-content terms. The release preserves the existing editing briefs and references; its changes concern packaging, readable case names, and explicit revision links. Source-video rights remain with their respective owners. Crayotter's software license applies separately.
Citation
If LongEditBench is useful for your research, please consider citing the dataset repository and the two Crayotter papers below. Please also acknowledge LongShOTBench for the original source-grounded interactions.
@misc{longeditbench2026,
author = {{Crayotter Team}},
title = {{LongEditBench}: A Benchmark for Long-Horizon Video Editing and Contextual Revision},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/datasets/witness0411/LongEditBench}},
url = {https://huggingface.co/datasets/witness0411/LongEditBench},
note = {Dataset, version 1.0; access by request}
}
@article{yan2026crayotter,
title = {Crayotter: Learning Long-Horizon Video Editing Agents via Group-Relative Preference Backpropagation},
author = {Yan, Lecheng and Lin, Jianze and Zhang, Yichong and Pan, Ben and Li, Wenxi and Lyu, Chenyang and Zhou, Liting and Gurrin, Cathal},
journal = {arXiv preprint arXiv:2608.02694},
year = {2026},
url = {https://arxiv.org/abs/2608.02694}
}
@article{yan2026crayottertraceable,
title = {Crayotter: Traceable Multi-Agent Workflows for Long-Form Video Editing},
author = {Yan, Lecheng and Zhang, Yichong and Xu, Xiantao and Lin, Jianze and Pan, Ben and Zheng, Xiaoyu and Qian, Jiawei and Wu, Anqi and Geng, Jiahui and Li, Ruizhe and Cai, Fengyu and Niu, Jingcheng and Li, Raymond and Li, Wenxi and Lyu, Chenyang},
journal = {arXiv preprint arXiv:2606.07636},
year = {2026},
url = {https://arxiv.org/abs/2606.07636}
}
- Downloads last month
- 84