GitHub
Filters harmful AI output by adjusting how the AI picks words, like a quiet safety guard.
Filters harmful AI output by adjusting how the AI picks words, like a quiet safety guard.
ReskLogits is a tool for making AI safer. It changes the raw scores that an AI creates when it's trying to decide what to say next. This helps prevent bad or unsafe content from being generated. It works behind the scenes so the user doesn't even notice, making it a "shadow ban" system.
Customizable Logits Processor.
ReskLogits lets you tweak how it works. You can tell it exactly which words or phrases to watch out for. This means you can make it fit your specific needs, whether you're a school or a bank.
Dynamic Content Filtering.
This tool filters content as it's being made. It doesn't wait until everything is done. This makes it really good at catching new harmful stuff right away and stops bad things from getting through.
Easy Integration with LLM Frameworks.
You can use ReskLogits with your current AI models. You don't need to change how your AI works or retrain it. This makes it super easy to add to what you're already using, saving you time and effort.
Comprehensive Logging and Monitoring.
ReskLogits keeps a close eye on everything. It records what it filters, and why. This helps you see how well it's working and if anyone is trying to get around its safety features.
Contextual Awareness in Filtering.
It's smart about filtering. It doesn't just block single words. It understands the whole sentence to avoid false alarms. So, good uses of a word won't get flagged as bad.
Integration with Existing Security Systems.
You can link ReskLogits with your other security tools. This creates a stronger defense
ReskLogits is an open-source Python library. It helps make large language models safer. It does this by filtering harmful content using something called a "shadow ban" system.
A shadow ban means that ReskLogits keeps the model from creating harmful content, but it does so subtly. Users won't get an explicit error message or a refusal. Instead, the model quietly steers away from the bad content. This makes the filtering less noticeable.
ReskLogits works by changing the raw scores (logits) that a language model uses to pick its next words. This helps prevent the model from generating bad content without completely blocking it or making the model's refusal obvious.
ReskLogits had its last release on November 15, 2025. This shows that the tool is being actively maintained and updated.
The Domain has been successfully submitted. We will contact you ASAP.
