Datasets:

Snooow1029
/

advbench-audio-inhouse

Tasks:

Audio Classification

Modalities:

Formats:

Languages:

Size:

Tags:

Libraries:

License:

Dataset card Data Studio Files Files and versions

xet

Community

Dataset Viewer

Auto-converted to Parquet Duplicate

Split (1)

train · 520 rows

Search is not available for this dataset

audio audioduration (s) 2.02 9.86

End of preview. Expand in Data Studio

AdvBench-Audio harmful (in-house TTS)

520 harmful AdvBench behaviors, single-speaker in-house TTS (neutral voice). Used as the SPIRIT noise-sensitive-neuron calibration source (clean vs random-noise activation diff). DISTINCT TTS from WeifeiJin/AdvBench-Audio (that repo ships advbench_audios/+AdvBench_Audio.json, the clean attack-source split loaded as advbench_clean). Layout mirrors dataset/advbench/.

Downloads last month: 61