Detoxify
github.com
- Free plan
- Linux
- Self-hosted
What Detoxify is
Detoxify provides trained models and code for classifying toxic comments from three Jigsaw challenges. It offers original, unbiased, and multilingual models based on BERT, RoBERTa, and XLM-RoBERTa. Depending on the model, outputs cover categories including threats, obscenity, insults, identity attacks, and explicit sexual content. The multilingual model is intended for English, French, Spanish, Italian, Portuguese, Turkish, and Russian. The project describes uses including research, fine-tuning on carefully constructed representative datasets, and helping moderators flag harmful content faster. It can be installed with pip and used from Python to predict labels for one string or a list of strings. Inference runs on CPU by default and accepts a PyTorch device such as CUDA. Inference uses Hugging Face Transformers and PyTorch Lightning; training also requires the Kaggle API to download data. The project warns that profanity-associated words may be classified as toxic regardless of tone or intent, potentially creating bias toward vulnerable minority groups. Detoxify is distributed under the Apache-2.0 license, and API access can be requested by email.
Who it suits
Detoxify suits researchers and developers building text moderation workflows who can use Python and account for the stated bias limitation. It may also help moderators flag harmful text for further attention.
Good to know, and what to watch
- Includes original, unbiased, and multilingual models.
- Supports seven languages with the multilingual model.
- Can predict labels for one string or a list.
- CPU inference is supported by default.
- Distributed under the Apache-2.0 license.
- Classifications may treat profanity as toxic regardless of intent.
- Does not moderate images, audio, or video.
- Does not include custom policies or human review workflow.
Plans
Every plan and priceOpen-source library
Free plan
The verdict
Detoxify offers a focused set of text-classification models that can support moderation and research. Its bias warning matters: profanity-related language can be flagged without regard to tone or intent.
What the maker publishes
- API access
- The repository says to email [email protected] to request access to its API. source
- Bias limitation
- The README warns that profanity-associated words may be classified as toxic regardless of tone or intent, potentially creating bias toward vulnerable minority groups. source
- Categories
- Model outputs include toxicity-related categories such as threats, obscenity, insults, identity attacks, and explicit sexual content, depending on the model. source
- Dependencies
- Inference uses Hugging Face Transformers and PyTorch Lightning, while training also requires the Kaggle API to download data. source
- Install
- The library can be installed with pip and used from Python to predict labels for one string or a list of strings. source
- Languages
- The multilingual model is intended for English, French, Spanish, Italian, Portuguese, Turkish, and Russian. source
- License
- The project is distributed under the Apache-2.0 license. source
- Models
- The library offers original, unbiased, and multilingual models based on BERT, RoBERTa, and XLM-RoBERTa, respectively. source
- Purpose
- Detoxify provides trained models and code to classify toxic comments from three Jigsaw challenges. source
- Runtime
- Inference supports CPU by default and accepts a PyTorch device such as CUDA. source
- Training data
- The models were trained for the 2018, 2019, and 2020 Jigsaw toxic-comment challenges using Wikipedia Comments and Civil Comments data. source
- Use
- The project describes its intended uses as research, fine-tuning on carefully constructed representative datasets, and helping moderators flag harmful content faster. source
Checked against the maker's pages on 8 Oct 2026.