Source code and details: https://git.anarchists.space/nemesis/Lingunymous
Everyone has a unique writing style, word choice, sentence structure, punctuation habits and that can be used to identify them. Lingunymous rewrites your text using a local LLM to mask these patterns, making it harder to attribute writing to a specific individual.
Use cases include:
- Protecting anonymity when publishing sensitive or controversial content
- Evading authorship attribution and stylometric analysis
- Separating your personal writing fingerprint from pseudonymous accounts
I’ve heard about this sort of technique being used, but an important step I usually hear about isn’t present here, which is to first translate the text to another language before translating it back. Simply asking a LLM to
seems likely to leak things like your word preferences and grammatical quirks, a 7b model isn’t going to be super creative about this and will want to take your lead on things. There needs to be an initial layer of stripping your statement of what makes it unique.