Kelly372/SemComp-Bench

SemComp

SemComp is an evaluation dataset for Semantic Task Completion Video Generation, an outcome-oriented video-generation setting in which a generated video is expected to achieve an instructed outcome while maintaining task-relevant semantic grounding with respect to a reference image.

Each instance provides metadata for reconstructing an image-instruction-video triplet:

• a reference image extracted from a full-context source video;
• paired brief and detailed instructions in English and Chinese;
• an outcome-centric video clip depicting successful task completion.

This repository distributes annotations, source URLs, and timestamps. It does not redistribute source videos or extracted frames.

Dataset Configurations

Configuration	File	Instances	Description
full	SemComp-Data.parquet	1,273	Complete SemComp evaluation dataset
core	SemComp-Core.parquet	60	Domain-balanced benchmark subset with 10 instances from each of six domains

Both configurations use the benchmark split because the released data is intended primarily for evaluation rather than model training.

SemComp covers six broad domains:

1. Arts and Precision
2. Beauty and Fashion
3. Crafts and DIY
4. Food and Cooking
5. Gardening and Pets
6. Sports and Fitness

Data Fields

Both Parquet files use the same schema.

Field	Type	Description
youtube_url	string	URL of the full-context source video
title	string	Title associated with the source video
domain	string	Broad task domain
category	string	Fine-grained task category
goal	string	Semantic goal or intended completed outcome
reference_timestamp	string	Timestamp of the reference frame in the full-context video
outcome_timestamp	string	Timestamp corresponding to the verified completed outcome
clip_start	string	Start timestamp of the outcome-centric clip
clip_end	string	End timestamp of the outcome-centric clip
clip_duration	float	Duration of the outcome-centric clip in seconds
brief_instruction_en	string	Brief English instruction
brief_instruction_zh	string	Brief Chinese instruction
detailed_instruction_en	string	Detailed English instruction
detailed_instruction_zh	string	Detailed Chinese instruction
reference_image_description_en	string	English description of the reference image
reference_image_description_zh	string	Chinese description of the reference image
align_type_label	string	Semantic alignment-type label
align_type_preserved_attributes	list[string]	Reference attributes expected to be preserved
align_type_raw_response	string	Raw response associated with alignment-type annotation
preserve_ignore_en	string	English summary of attributes to preserve or ignore
preserve_ignore_zh	string	Chinese summary of attributes to preserve or ignore

Loading the Dataset

Install the Hugging Face Datasets library:

pip install datasets

Load the complete dataset:

from datasets import load_dataset

dataset = load_dataset(
    "Kelly372/SemComp",
    "full",
    split="benchmark",
)

Load the SemComp-Core benchmark subset:

from datasets import load_dataset

core = load_dataset(
    "Kelly372/SemComp",
    "core",
    split="benchmark",
)

Because full is the default configuration, it can also be loaded without specifying a configuration:

dataset = load_dataset(
    "Kelly372/SemComp",
    split="benchmark",
)

The dataset repository is hosted under the Kelly372 Hugging Face namespace.

Accessing the Visual Data

Each row contains the metadata needed to locate the corresponding visual material:

1. youtube_url identifies the full-context source video.
2. reference_timestamp identifies the reference frame.
3. clip_start and clip_end identify the outcome-centric segment.
4. outcome_timestamp identifies the verified completed outcome.

Users are responsible for accessing source media only where permitted by applicable law, platform terms, geographic restrictions, and rights-holder permissions.

Source videos may be removed, made private, restricted by region, or otherwise become unavailable over time.

Intended Uses

SemComp is intended for non-commercial research involving:

• semantic task-completion video generation;
• image-to-video and text-to-video evaluation;
• reference-grounded video generation;
• outcome-achievement evaluation;
• generation-reliability evaluation;
• instruction-specificity analysis;
• multilingual instruction analysis;
• video-generation benchmark development.

SemComp-Core is the recommended subset for standardized model comparisons. SemComp-Data is appropriate for broader analysis across domains and task categories.

The suitability of this dataset for training video-generation models has not been established.

Limitations

• The repository contains metadata and annotations rather than redistributed video files.
• Some source URLs may become unavailable after release.
• Access to source videos may vary by region or platform policy.
• The dataset covers six broad domains but does not represent every real-world task.
• Automatically assisted annotations and instructions may contain errors or ambiguities.
• English and Chinese instructions are intended to describe the same outcome but may contain minor linguistic differences.
• The domain and category distributions should not be interpreted as representative of all online videos.
• Performance on SemComp-Core should not be interpreted as a complete measure of general video-generation capability.
• The presence of a source URL does not imply that the source media is licensed for redistribution, commercial use, or model training.

License

SemComp is released under the SemComp Non-Commercial Research License. See LICENSE for the complete terms.

SemComp metadata and annotations are available for non-commercial research, study, instruction, testing, and validation only. No commercial license is granted.

SemComp uses metadata and source references from Koala-36M. Users must also comply with the Koala-36M license.

The source videos are not redistributed in this repository. They remain subject to the rights of their respective owners, licenses associated with their original sources, and applicable platform terms.

Citation

If you use SemComp or SemComp-Core, please cite:

@article{semcompbench2027,
  title   = {SemComp-Bench: Benchmarking Semantic Task Completion in Video Generation},
  author  = {Anonymous Authors},
  journal = {arXiv preprint},
  year    = {2027}
}

Replace the anonymous author entry with the final author list after the review process.

Contact

For questions, corrections, or reports of unavailable or inappropriate source links, please open a discussion in the dataset repository's Community tab.

Version

Initial release: v1.0.0