Verus-Bench / README.md
SuperCarryZZR's picture
Release Verus-Bench
35dc3e8 verified
|
Raw History Blame Contribute Delete
4.71 kB
metadata
license: other
license_name: verus-noncommercial-research-evaluation
license_link: LICENSE
pretty_name: Verus-Bench
task_categories:
  - video-classification
tags:
  - video
  - benchmark
  - artifact-detection
configs:
  - config_name: default
    data_files:
      - split: test
        path: metadata/test.parquet

Verus-Bench

Verus-Bench is a video artifact evaluation dataset from Verus-Team, containing 192 test samples: 96 generated (FAKE) and 96 natural (REAL), with 320 human-annotated artifact findings. GT includes labels, artifact categories, descriptions, temporal intervals and spatial boxes. There is a single test split.

Contents

Content Included in this repository Complete benchmark
Samples and human GT 192 192
Frame images 5,204 across 166 samples 5,932 across 192 samples
Source videos 142 168 available; 24 samples have frames only

The samples comprise 96 HJY generated videos, 24 HJY natural samples, and 24 samples each from NExT-QA, OneThinker/DiDeMo and OneThinker/STVG. For 26 third-party samples, media is omitted under the distribution policy: 14 have ND licenses, 9 have unresolved licenses and 3 have incomplete historical attribution. Their GT, source references and frame metadata are included. Follow SETUP_ALL_DATA.md to obtain and prepare the missing inputs.

File or directory Purpose
frames/ Benchmark input images, grouped by sample ID
videos/ Optional source videos for viewing or preprocessing
ground_truth.jsonl Human test labels and findings
metadata/test.parquet Default Hugging Face test table, including GT and media paths
metadata/frames.jsonl Frame order, timestamps, dimensions and extraction settings
metadata/inputs.jsonl GT-free input index; omitted media has null image paths
sample_ids.txt, metadata/samples.csv Sample membership and compact overview
THIRD_PARTY_DATA.md, THIRD_PARTY_DATA.csv Source attribution and media licenses
metadata/reference_media.csv The 26 samples requiring locally obtained media
RELEASE_MANIFEST.json Dataset version and counts

Use the data

Model inference uses the environment in the code repository. To read the dataset metadata, install Hugging Face Datasets with python -m pip install datasets:

from datasets import load_dataset

test = load_dataset("Verus-Team/Verus-Bench", split="test")

The Parquet table stores relative media paths, so download the repository's frames/ directory as well. Paths resolve from the dataset root. The table contains GT; model inputs must use only the images and timestamps. See SETUP_ALL_DATA.md to prepare a complete GT-free metadata/inputs.local.jsonl. To run only the supplied media, explicitly export metadata/inputs.available.jsonl with --available-only; this is a 166/192 partial run. The generated manifests work directly with the companion code's --input-jsonl option.

Use the supplied frames in their recorded order. Source videos may cover a longer interval than the benchmark sample; source_window_seconds records the first and last sampling-grid times in the upstream file. Its second value is not an exclusive clip end. Audio is not an input.

The GT supports artifact detection, category evaluation, temporal localization and spatial localization. Boxes use normalized XYXY coordinates in [0,1]. REAL samples have empty findings. See docs/DATA_SCHEMA.md for field definitions.

Licensing

Verus-owned material is for non-commercial research and evaluation only; see LICENSE. Third-party media and its frames retain their individual original licenses, including attribution, NC and SA conditions where applicable. See THIRD_PARTY_DATA.md and the per-item CSV for terms and available attribution.

Coverage and reconstruction

A complete benchmark run uses all 192 samples. Report the evaluated IDs and coverage for partial runs. Upstream sources may become unavailable. Re-extraction follows the recorded sampling settings; decoder and encoder differences can change the resulting images. File presence and frame counts indicate that inputs are available, not that reconstructed images are identical to the supplied frames.