ChatSensitiveWords is an open-source project that uses an LLM and a sensitive word lexicon to automatically determine sensitive words.
Purpose The project targets scenarios that require sensitive word detection for text content. According to the materials, it mainly targets politically sensitive words and pornographic sensitive words. The project team states that the project has been launched on chatpaper because chatpaper freely provides the full functionality of the academic version of GPT, which includes a hidden chat feature. They originally hoped to provide an academic chat interface, but in order to respond to user curiosity or deliberate attacks on output by a very small number of people, they needed to urgently launch sensitive word detection.
Main capabilities The project uses a combination of an LLM and a sensitive word lexicon to automatically determine whether text involves sensitive words. The materials mention that if there are sensitive words, Chat is asked to judge the semantics to confirm whether there is a false positive. In other words, it does not rely only on the lexicon for direct blocking, but adds a model semantic judgment step. The project also discloses performance: when there are no sensitive words, detection takes about 0.2 seconds; when there are sensitive words, it takes about 0.2+0.8 seconds.
How to use The project is provided as a code repository on GitHub, and the repository address is https://github.com/kaixindelele/ChatSensitiveWords/。资料给出的安装命令为：pip install -r requirements.txt. The project team welcomes everyone to test and evaluate locally, and specifically reminds people: "Do not test on our academic version." Apart from that, the materials do not provide more detailed usage examples, interface descriptions, or configuration steps.
Applicable scenarios From the project description, it is suitable for scenarios that need to add sensitive word detection capability to conversations or output content, especially products like an academic chat interface that want to prevent sensitive output while allowing open conversation. The project team also mentions that if it is helpful to friends building websites, donations are welcome, and requests that it not be packaged and sold for money.
Limitations and current status The project team clearly states that traditional sensitive word lexicon detection has not yet been done well, and then the Chat prompt has also not been tuned to optimal. This means its detection effectiveness and stability may still have room for improvement. The materials do not provide more details such as accuracy, false positive rate, supported languages, deployment environment requirements, or license information. The project welcomes users to star and pull new features, and asks that the repository be cited when using the data or code in the repository.
Other notes The materials do not provide information about the project's price, charging model, or commercial version; nor do they provide specific usage methods, API documentation, or interface descriptions other than pip installation.
