OmniClean

OmniClean is a leakage-aware omni-modal evaluation set built from retained examples across 9 source benchmarks. It is designed to reduce visual-shortcut effects in omni evaluation by applying visual-only probing where query-level filtering is defined, while keeping selected full subsets for protocol-exception benchmarks where a filtered subset is undefined or intentionally not reported.

This release contains 8,551 evaluation examples in a minimal slim JSONL format.

What this release is

Raw omni benchmark scores can be inflated by visually answerable examples. OmniClean is intended to provide a cleaner evaluation target for audio-visual-language QA and related omni understanding tasks.

This release is for evaluation. It is not intended as a training corpus.

Composition

Total examples: 8,551

Source benchmark (dataset_source)	Examples	Notes
AV_Odyssey_Bench	4555	Full selected subset retained as a protocol exception
VideoHolmes	885	Query-level cleaned subset
WorldSense	875	Query-level cleaned subset
IntentBench	660	Query-level cleaned subset
OmniBench	417	Query-level cleaned subset
CG-AV-Counting	376	Full selected subset retained as a protocol exception
OmniVideoBench	318	Query-level cleaned subset
Daily-Omni	237	Query-level cleaned subset
UNO-Bench	228	Query-level cleaned subset

Data format

Each record contains the following fields:

• dataset_source: source benchmark name
• source_id: source sample identifier
• question: question text
• options: candidate answers; may be empty for some benchmarks
• answer: benchmark-native gold answer
• media_paths: relative media references with image, audio, and video lists
• question_type: benchmark-native question category; may be null

Example:

{
  "dataset_source": "OmniVideoBench",
  "source_id": "omnivideobench:0",
  "question": "Before picking up the kitten, the blogger explains a sign. Which concepts can it be associated with?",
  "options": [
    "A.Ancient Chinese stories and Japanese anime",
    "B.Ancient Chinese Imperial Palace Architecture and Japanese Bar Names",
    "C.A certain type of Chinese cuisine and a certain type of Southeast Asian opera",
    "D.Chinese garden art and Western palace architecture"
  ],
  "answer": "Ancient Chinese stories and Japanese anime",
  "media_paths": {
    "image": [],
    "audio": [],
    "video": ["videos/video_1.mp4"]
  },
  "question_type": "reference reasoning"
}

Important notes

Benchmark-native answers

answer is not normalized into a single format across all sources. Depending on the benchmark, it may be:

• a single option letter such as A
• multiple option letters such as D,E,F
• a numeric answer such as 18
• the full answer text
• a short free-form label such as Yes

Evaluation should therefore use benchmark-aware answer normalization.

Optional fields by source benchmark

• options can be empty for some examples.
• question_type can be null for some examples.
• media_paths always contains the keys image, audio, and video, but some lists are empty.

Protocol exceptions

Two source benchmarks are intentionally retained as selected full subsets in this release:

• AV_Odyssey_Bench: a visual-only filtered subset is not defined because some answer options contain audio-bearing content.
• CG-AV-Counting: visual-only probing is used diagnostically, but a filtered-score benchmark is not reported because further exclusion would overly shrink an already difficult subset.

Loading with datasets

from datasets import load_dataset

ds = load_dataset("che111/OmniClean", "slim", split="test")
print(ds[0])

Limitations

• This release keeps benchmark-native answer formats instead of forcing a single unified answer schema.
• Source benchmarks differ in modality structure: some examples are video-only, some are image+audio, and some are audio+video.
• Relative paths in media_paths should be interpreted with respect to the released data layout.

Citation

If you use OmniClean, please cite the accompanying paper:

@misc{omniclean2026,
  title={Probing and Boosting Omni Understanding: Leakage-Aware Evaluation and a Staged Post-Training Study},
  author={StepFun-Audio Team},
  year={2026}
}

License

Please replace this section with the final license and confirm that redistribution terms are compatible with all included source benchmarks and media assets.