OST-DiagBench

OST-DiagBench contains exactly four diagnostic operations with 510 cases per operation: 2,040 cases in total. Every case includes an MP4 input and a paired M4A audio file.

Operation	IDs	Construction	What it tests
MUTE	00001–00510	Keep the video unchanged and remove the visually suggested source sound from the soundtrack; non-target background audio may remain.	Whether a model claims to hear a sound suggested by the video even when that sound is absent.
SWAP	00511–01020	Keep the video unchanged, remove the original source sound, and insert an unrelated donor sound.	Whether a model follows the sound that is actually audible instead of the visual suggestion.
MIX	01021–01530	Keep the video unchanged, retain the original source audio, and add a donor sound.	Whether a model recognizes both simultaneously audible source and donor sounds.
CLASH	01531–02040	Keep the on-screen subtitle/visual fact, but replace the corresponding spoken fact with a mutually exclusive value.	Whether a model answers from the spoken audio or copies the conflicting on-screen subtitle.

Files

data/
  mute.jsonl             510 MUTE rows
  swap.jsonl             510 SWAP rows
  mix.jsonl              510 MIX rows
  clash.jsonl            510 CLASH rows
  audio_event.jsonl      MUTE + SWAP + MIX (1,530 rows)
  ost_diagbench.jsonl    all four operations (2,040 rows)
media/{mute,swap,mix,clash}/
  0001.mp4 ... 0510.mp4
  0001.m4a ... 0510.m4a
eval/                    scoring scripts

The public id is unique across the full dataset. operation_local_id runs from 0001 to 0510 inside each operation. See data/SCHEMA.md for the manifest fields and LICENSES.md for upstream media terms.

Download

pip install -U huggingface_hub
hf auth login
python - <<'PY'
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="EnjunDu/OST-Diagbench",
    repo_type="dataset",
    local_dir="OST-DiagBench",
)
PY

Install the evaluation dependencies from the downloaded directory:

cd OST-DiagBench
pip install -r requirements.txt

MUTE, SWAP, and MIX evaluation

Run each MP4 in data/audio_event.jsonl with the exact prompt:

Describe all salient sounds you hear in this clip.

Write one prediction per line:

{"id":"00001","description":"I hear wind and distant traffic.","model":"my-model"}

Then score the predictions:

python eval/evaluate.py \
  --data data/audio_event.jsonl \
  --predictions predictions_audio_event.jsonl \
  --output-dir outputs/audio_event

The scorer reports MUTE visual-hallucination recall, SWAP audible-donor and visually suggested source recall, and MIX source/donor/joint recall.

CLASH evaluation

For each CLASH row, ask its question twice:

1. condition="av": run media_path (the MP4 with conflicting speech and subtitle).
2. condition="audio_only": run audio_path (the paired M4A without the visual subtitle).

Write predictions without copying the ground-truth fields into the prediction file:

{"id":"01531","condition":"av","answer":"64 percent","model":"my-model"}
{"id":"01531","condition":"audio_only","answer":"64 percent","model":"my-model"}

Score them by joining against the CLASH manifest:

python eval/score_clash.py \
  --data data/clash.jsonl \
  --predictions predictions_clash.jsonl \
  --output outputs/clash_scored.jsonl

Each answer is labeled audio, subtitle, both, or neither, with separate summaries for the AV and audio-only conditions.