MuTox annotations

This repo contains the annotations from MuTox, Meta's multilingual audio-based toxicity detection benchmark covering 30 languages.

What's here

The original mutox.tsv annotations file, unmodified except for being split by its partition column into four files/splits:

split	file	rows
train	mutox_train.tsv	67,971
dev	mutox_dev.tsv	22,273
devtest	mutox_devtest.tsv	10,691
unpartitioned	mutox_unpartitioned.tsv	4,561

devtest is the split used for benchmarking (e.g. by the inspect-evals-mutox eval). The unpartitioned rows have an empty partition value in the source file and are mostly contains_toxicity = "Cannot say" rows that Meta excluded from all three named partitions; they're included here for completeness rather than because they belong in any benchmark split.

Columns:

column	description
id	sample id
lang	MuTox language code (e.g. eng, arb; note Western Persian is pes, not fas)
partition	train, dev, devtest, or empty (see split table above)
public_url_segment	"<source url> <start offset> <end offset>" identifying the audio span. Offsets are milliseconds for eng/spa and 16 kHz sample counts for the other 28 languages
audio_file_transcript	transcript of the utterance
contains_toxicity	annotator-facing label (Yes/No)
toxicity_types, perlocutionary_effects	annotation detail fields
label	binary toxicity label (1/0) used as the ground truth
etox_result, detoxify_score, mutox_speech_score, mutox_text_score, mutox_zero_shot_speech_score, mutox_zero_shot_text_score	scores from Meta's reference toxicity classifiers

What's not here: audio

No audio is redistributed in this dataset. public_url_segment points at third-party sources (podcast CDNs, archive.org, etc.) that MuTox indexes but does not own, so redistributing clipped audio here would exceed the rights granted by the annotations license below. Source URLs also rot over time, so the number of clips you can actually materialize will be lower than the row count.

To build the audio locally, fetch each public_url_segment URL and cut out [start, end) (converting units per the language, see table above). A reference implementation that does this fetching/clipping/caching, wired up as an Inspect AI eval, is at ymerkli/inspect-evals-mutox (src/mutox/dataset.py).

License

The MuTox annotations are MIT licensed by Meta Platforms, Inc. — see LICENSE. The underlying audio remains the property of its original third-party hosts and is not covered by this license.

Citation

@article{mutox,
  title={MuTox: Universal MUltilingual Audio-based TOXicity Dataset and Zero-shot Detector},
  author={Costa-juss{\`a}, Marta R. and Meglioli, Mariano Coria and Andrews, Pierre and Dale, David and
  Hansanti, Prangthip and Kalbassi, Elahe and Mourachko, Alex and Ropers, Christophe and Wood, Carleigh},
  journal={arXiv preprint arXiv:2401.05060},
  year={2024}
}