Detoxify
UnitaryDetoxify is a Python library and a set of trained models that score text for toxicity, insults, threats, obscenity and identity attacks. Its three models were trained on the Jigsaw toxic-comment challenges, and the multilingual model covers seven languages. Unitary built it, and the repository also holds the training code and configurations.
Openness
5 high confidence- license
- Apache-2.0 (OSI, the code and the default checkpoints published in the repository's releases)
- source
- public (library, training script and configurations; default checkpoints download from the repository's own releases)
- core features withheld
- no — no enterprise path in the repository tree; Unitary's commercial product is an unrelated automation service
Detoxify's code is Apache-2.0, and the checkpoints it loads by default are published in the same repository's releases with no separate license. Unitary's commercial business is an unrelated automation service, and nothing in the library is held back for it.
- https://pypi.org/pypi/detoxify/json recorded 2026-11-12
The PyPI record for detoxify 0.5.4, uploaded 2026-10-14, carries the classifier "License :: OSI Approved :: Apache Software License" and the homepage https://github.com/unitaryai/detoxify.
- https://raw.githubusercontent.com/unitaryai/detoxify/HEAD/LICENSE recorded 2026-11-12
The LICENSE body is the stock Apache License 2.0 text with no added condition.
- https://raw.githubusercontent.com/unitaryai/detoxify/HEAD/README.md recorded 2026-11-12
The README installs with "pip install detoxify", documents training with "python train.py --config ..." for each challenge, and says checkpoints "can be downloaded from the latest release or via the Pytorch hub API". Searched for enterprise, pricing, paid and license key: none.
- https://raw.githubusercontent.com/unitaryai/detoxify/master/detoxify/detoxify.py recorded 2026-11-12
The library sets DOWNLOAD_URL = "https://github.com/unitaryai/detoxify/releases/download/" and loads its default checkpoints from the repository's own releases.
- https://ungh.cc/repos/unitaryai/detoxify/files/master recorded 2026-11-12
The repository file listing shows the detoxify package, train.py, hubconf.py and configs, and no ee, enterprise, commercial or proprietary path.
- https://www.unitary.ai/ recorded 2026-11-12
Unitary's homepage sells "Virtual Agents" that "automate manual work across insurance, financial services, healthcare" and does not mention Detoxify or toxicity classification.
Adoption
2 high confidencePyPI downloads of the detoxify package, which is how the library and its default models are installed. Loading the checkpoints straight from the releases page or through Torch Hub skips the package and is not counted.
- https://pypi.org/pypi/detoxify/json recorded 2026-11-12
The package's only project URL is https://github.com/unitaryai/detoxify, the product's repository.
- https://pypistats.org/api/packages/detoxify/recent recorded 2026-11-12
"last_month":62494 for detoxify
Capability
2 medium confidenceDetoxify scores one kind of harm, toxic language, under seven labels, and runs on a CPU by default. It does not judge prompts or replies against a policy or cover hazards such as violent crime or self-harm, which Llama Guard screens across fourteen categories, so it sits two steps below it.
- https://raw.githubusercontent.com/unitaryai/detoxify/HEAD/README.md recorded 2026-11-12
The README's results table gives Detoxify scores of 98.64, 93.74 and 92.11 for the original, unbiased and multilingual models, lists the labels toxicity, severe_toxicity, obscene, threat, insult, identity_attack and sexual_explicit, names bert-base-uncased, roberta-base and xlm-roberta-base as the base models, and says the multilingual model covers seven languages.
Verified 2026-11-12