lynote-ai/humanize-text-zh

humanize-text-zh is an open source Python toolkit, written mostly in Chinese, that rewrites AI generated text so it reads as if a person wrote it. The README presents it as a research and learning project for people who want to see how AI humanization techniques actually work, and it ships four separate methods rather than a single hidden pipeline. The same authors also run a paid web product called Lynote.ai, and the README is upfront that the open source code is the building blocks while the hosted service mixes them together for end users. The first method is a chain of machine translations through distant languages, for example English to Chinese to Japanese to Finnish and back to English. The structural differences between those languages tend to reshape sentences. The README lists Google Translate, Niutrans, MyMemory and Apertium as the engines it can call, and notes that very long academic text can lose technical terms when too many translation hops are used. The second method is multi round rewriting with a large language model. It uses the DeepSeek API at a high temperature (1.1 to 1.3) and a burst oriented prompt to vary sentence length and word choice over two or three passes. The README warns that each extra pass can drift further from the original meaning. The third method is a closed loop where the text is rewritten, then scored by four AI text detectors at once (Binoculars on GPT-2, a RoBERTa classifier, statistical features, and diversity metrics), and any paragraph that still trips a detector is rewritten again. Document level rewriting feeds into sentence level rewriting and then into rule based cleanup, including replacing about thirty common AI English words and eleven common Chinese AI phrases. The README says this method needs local detection models and a GPU. The fourth method blends outputs from several neural translation engines in a single pass to avoid any single model's fingerprint. The project ships under the MIT license with a pip install path and a docker compose path that exposes an HTTP API on port 8000.