
from humanize-chinese145
Detects and humanizes Chinese AI-generated text with rule-based and statistical features; offers CLI modes for detection, rewriting and academic de-AIGC workflo
Humanize Chinese is a local, zero-dependency toolset and prompt-guide for detecting and rewriting Chinese AI-generated text. It combines 20+ rule-based detection categories with 8 HC3-calibrated statistical features (sentence-length CV, short-sentence fraction, comma density, GLTR buckets, surprisal metrics) to score AIGC likelihood and offers adaptive rewrite strategies and style transforms (casual, Zhihu, Xiaohongshu, WeChat, academic). The project provides a unified CLI for detect/rewrite/academic/style/compare workflows and extensive templates for paraphrase and academic replacements.
Use when a user asks to "remove AI flavor", "humanize text", "reduce AIGC score", or to prepare academic content for submission (CNKI/VIP/Wanfang) that needs AIGC reduction. Appropriate for editors, researchers, and platforms needing local, auditable AIGC mitigation.
Inferred compatibility: code-capable agents and local execution environments (agents with Read/Write/Exec privileges, Claude Code or Copilot-like agents).
Chinese AI-text detection and humanization toolkit with 24 scripts including unified CLI, detection, rewriting, style transforms, and LR model training. Core CLI (humanize.py) runs cleanly and shows well-documented help. Several scripts fail gracefully on empty input (expected in DRY_RUN). Training scripts hardcode local macOS paths and need external corpora. Checked-in JSON data assets are 0B/missing (noted by check_assets.py doctor). No security concerns — pure Python, offline, no network calls, no credentials.
Impressive niche tool — well-documented, no dependencies, offline-first. Main issue is missing data assets in the repo (0B JSON files) which means detection accuracy degrades. SKILL.md is thorough but monolithic (no references/ dir for progressive disclosure). Training scripts are included but require external corpora not in the repo.