lynote-ai/humanize-text

What it solves

This project provides a pipeline for rewriting AI-generated text to make it sound more natural and human-like, specifically aiming to break the statistical fingerprints that AI detectors typically identify.

How it works

The project offers two distinct approaches:

1. Structure-First Data Pipeline: The current primary method focuses on creating high-quality training data for a custom rewrite model. It uses Supervised Fine-Tuning (SFT) and Kahneman-Tversky Optimization (KTO) to teach a model to rewrite the actual narrative structure of text rather than just changing surface-level words.
2. Translation Chain (Legacy): A runnable reference pipeline that uses a multi-hop translation process. It rewrites text using LLMs (translating to Chinese and then Japanese) and then passes it through multiple translation engines (Google Translate and Niutrans) to Finnish and then back to English to maximize linguistic distance and structural disruption.

Who it’s for

Developers and writers who want to improve the readability and natural cadence of AI-assisted drafts or explore the methods used to bypass AI detection.

Highlights

• Multi-Tiered Approach: Offers Standard, Advanced, and Focus tiers with varying levels of restructuring depth.
• SFT + KTO Training: Implements a preference-based training pipeline to prioritize genuine structural changes over simple edits.
• Multi-Engine Translation: Uses a combination of LLMs and NMT engines to eliminate single-engine fingerprints.
• Deterministic and Auditable: Includes SHA-256 identity per pair and versioned YAML policies for data construction.