VisualArtifactSpec / README.md
NaturalKnight's picture
Add arXiv report and citation
ffae989 verified
|
Raw History Blame Contribute Delete
6.88 kB
metadata
pretty_name: VisualArtifactSpec 1.0
license: cc-by-4.0
language:
  - en
  - zh
size_categories:
  - n<1K
tags:
  - benchmark
  - visual-artifacts
  - agent-evaluation
  - rubric-evaluation
configs:
  - config_name: default
    data_files:
      - split: test
        path: tasks.jsonl

VisualArtifactSpec 1.0

Evaluating Visual Artifacts under Task Specifications

Website · Report · GitHub

211 tasks · 4 artifact families · 14 task types · 8 application purposes

VisualArtifactSpec (VAS) evaluates how agents turn source materials and task specifications into visual artifacts that serve practical needs. Tasks span creation and modification, with explicit objectives, materials, constraints, and deliverables. Evaluation separates Task Fulfillment from Artifact Performance, assessing both requirement satisfaction and suitability for the intended audience, purpose, and medium.

VAS task families and illustrative examples

Illustrations show task coverage, not model results.

Task coverage

Artifact family Tasks Task types
Presentations 47 Original decks, template adaptation, existing-deck revision
Static Graphics 72 Posters, infographics and visual summaries, banners, multi-page series
Animations & Videos 32 Animated graphics, animated videos, edited videos
Web Interfaces 60 Content browsing, data exploration, tools and simulations, stateful workflows

Categories follow the primary authoring objective rather than the file extension. Each task includes instructions, input attachments, evaluator guidance, and task-specific rubrics. The dataset contains 11,718 rubric items: 7,920 requirement checks and 3,798 quality items, with 3,125 acceptance groups.

Download

Browse the tasks or download the complete dataset below.

pip install huggingface_hub datasets
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Accio-Lab/VisualArtifactSpec",
    repo_type="dataset",
    local_dir="vas-bench",
)

Attachments are included under each task's inputs/ directory. Pass vas-bench/data to the local Docker Judge and scoring tools. Generate submissions using your own model and harness.

Dataset layout and programmatic access
data/
  benchmark.json
  tasks/TASK_ID/
    prompt.md
    platform_instructions.txt
    rubric.json
    evaluation_prompt.txt
    task.json
    inputs.manifest.json
    inputs/
tasks.jsonl
  • prompt.md and platform_instructions.txt: agent-facing task and execution requirements.
  • rubric.json and evaluation_prompt.txt: evaluator criteria, answer anchors, weights, and instructions.
  • task.json: task metadata, acceptance groups, and objective definitions.
  • inputs.manifest.json: attachment paths and hashes.

Task IDs use category names and three-digit numbers, such as presentations-001 and web-interfaces-001. Attachment filenames and nested paths follow the task's references.

The dataset viewer and Python API read tasks.jsonl:

from datasets import load_dataset

tasks = load_dataset("Accio-Lab/VisualArtifactSpec", split="test")

Fields ending in _json contain JSON strings. Download the repository snapshot to obtain the actual input files. Pin the dataset commit when reporting or comparing evaluations.

Evaluation

Task Fulfillment and Artifact Performance rubric

Task Fulfillment establishes the base score (0–50). Compliance checks specified obligations; Functionality checks required behaviors where applicable. Each acceptance group represents one independently acceptable output or behavior. All requirements in a group must pass, and groups contribute equally to the shared 50-point pool.

Artifact Performance provides an adjustment (−50 to +50). Eighteen shared criteria assess Content, Organization, Visuals, and Delivery, using task-specific anchors. Ratings from −3 to +3 are mapped to points before weighted aggregation.

Final score = fulfillment base + performance adjustment

Range: −50 to 100 · Pass: 50 or above · Negative totals are retained.

Judgments must cite inspectable evidence from source materials, rendered views, native structures, temporal observations, or recorded interactions. Unresolved evaluations remain unscored.

Scoring details
Criterion rating −3 −2 −1 0 +1 +2 +3
Mapped points −50 −25 −10 0 +10 +25 +50

Average mapped points within each dimension using rubric item weights, then combine Content (25%), Organization (15%), Visuals (50%), and Delivery (10%).

B = 50 × passed acceptance groups / all acceptance groups
Δ = 0.25 C + 0.15 O + 0.50 V + 0.10 D
S = B + Δ

C/O/V/D are dimension point scores after mapping, not raw ratings. Zero denotes ordinary quality without substantive defects. Positive levels require the progressively stronger evidence specified by each item's anchors.

Keep source materials and submitted artifacts unchanged. Retain item-level judgments and evidence, report completion coverage and evaluation failures, and compare common tasks under matching rubrics, grouping, and scoring rules. Complete saved judgments can be rescored with the code repository's scoring tools.

For observation requirements and rating boundaries, see the Judge protocol.

Citation

If you use VisualArtifactSpec in your research, please cite our report:

@misc{ouyang2026visualartifactspec10evaluatingvisual,
  title={VisualArtifactSpec 1.0: Evaluating Visual Artifacts under Task Specifications},
  author={Shuyi Ouyang and Xiaojun Tang and Zijian Zou and Xintong Zhang and Mingli Song and Guannan Zhang and Hongwei Xue},
  year={2026},
  eprint={2611.21382},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2611.21382},
}

License

Original task instructions, rubrics, and illustrations use CC BY 4.0; the code and Judge harness use Apache-2.0; third-party attachments and components retain their original licenses and attribution requirements.