sgaofen/humanizer-local-model

What it solves

This project provides a specialized 12B parameter model designed to rewrite AI-generated text (such as emails, essays, and reports) so that it sounds more natural and human-like. It specifically addresses the common "robotic" or overly formal tone of LLM outputs while maintaining strict factual fidelity to the original draft, including numbers, dates, names, and units.

How it works

The model is a fine-tuned version of Google's Gemma-4-12B. It was trained using a combination of supervised fine-tuning on human-written text, Direct Preference Optimization (DPO), and Reinforcement Learning (GRPO) to optimize for fact fidelity and natural phrasing. Crucially, no AI detection tools were used during the training process. The project includes a cross-platform desktop application that handles model downloading and local inference, allowing users to process text offline.

Who it’s for

It is for professionals, students, and writers who use AI to draft content but need to ensure the final output is indistinguishable from human writing and factually accurate. It is also useful for developers who want to integrate a "humanizing" step into their own AI agent workflows.

Highlights

• Runs entirely offline on your local machine (macOS and Windows apps available).
• Supports both English and Chinese text.
• Includes a command-line tool (hz) for batch processing long documents like Markdown or Word files.
• Optimized with quantization-aware training to maintain high performance even in smaller, compressed file sizes.
• Designed to preserve formatting, headings, and code blocks during the rewriting process.