MuTox: Multilingual Audio-based Toxicity Detection Benchmark

MuTox evaluates a model’s ability to detect toxicity in short speech clips across 30 languages. The dataset contains human-annotated audio utterances (20k for English/Spanish, 4k for 28 other languages) sourced from CommonVoice and SeamlessAlign, pre-selected using DETOXIFY and ETOX text classifiers. The inspect eval presents audio clips and asks the model to classify toxicity via binary yes/no response, scored with quasi-exact match against human-annotated labels.

Maintained By

@ymerkli

Code

mutox

Paper

https://arxiv.org/abs/2401.05060

Overview

⚠️ External evaluation. Code lives in an upstream repository. inspect_evals lists it for discoverability; review the upstream repo and pinned commit before running.

Source: ymerkli/inspect-evals-mutox@b5913be

MuTox evaluates a model’s ability to detect toxicity in short speech clips across 30 languages. The dataset contains human-annotated audio utterances (20k for English/Spanish, 4k for 28 other languages) sourced from CommonVoice and SeamlessAlign, pre-selected using DETOXIFY and ETOX text classifiers. The inspect eval presents audio clips and asks the model to classify toxicity via binary yes/no response, scored with quasi-exact match against human-annotated labels.

Usage

Installation

This is an externally-maintained evaluation. Clone the upstream repository at the pinned commit and install its dependencies:

git clone https://github.com/ymerkli/inspect-evals-mutox
cd inspect-evals-mutox
git checkout b5913be2e31f41563c761f86864f124ace5e1a78
uv sync

Running evaluations

CLI

uv run inspect eval src/mutox/mutox.py@mutox --model openai/gpt-5-nano

Python

from inspect_ai import eval
from mutox.mutox import mutox

eval(mutox(), model="openai/gpt-5-nano")

View logs

Log viewing (inspect view) and default-model setup are documented in the Inspect Evals README.

More information

For the dataset, scorer, task parameters, and validation, see the upstream repo: ymerkli/inspect-evals-mutox.

Options

You can control a variety of options from the command line. For example:

uv run inspect eval src/mutox/mutox.py@mutox --limit 10 --sample-shuffle
uv run inspect eval src/mutox/mutox.py@mutox --max-connections 10
uv run inspect eval src/mutox/mutox.py@mutox --temperature 0.5

See uv run inspect eval --help for all available options.

More command-line options: Inspect docs ↗

Static checks

Results of inspect-evals-lint at the registered commit. Rule names link to their documentation; findings link to the file at that commit. These checks describe structure and conventions, not whether the evaluation measures what it claims.

Loading lint results…