gohumanize/gohumanize-open-humanizer

GoHumanize Open Humanizer Overview

The GoHumanize Open Humanizer is a 4 billion parameter model, fine-tuned from Qwen3-4B using QLoRA, specifically designed to transform AI-generated English text into more natural, human-like prose. Developed by the GoHumanize team, this model was trained on a dataset of 2,000 pairs of AI-styled passages and human-written originals sourced from Project Gutenberg.

Key Capabilities

• Text Humanization: Rewrites AI-styled English to incorporate varied sentence length, concrete wording, and natural rhythm, while preserving original meaning and facts.
• Research and Education: Serves as a public model to illustrate the methodology behind AI text humanization systems, including dataset preparation, fine-tuning, and evaluation.
• Performance: Demonstrates improved BERTScore F1 (0.921 vs 0.900 for base Qwen3-4B) and ROUGE-L (0.540 vs 0.424) against human targets, indicating closer meaning and wording.
• Efficiency: Achieves a length ratio closer to human text (0.928 vs 0.830 for base) and significantly reduces stock LLM phrases.

Good for

• Developers and Researchers: Ideal for understanding and experimenting with text humanization techniques.
• Educational Purposes: Useful for demonstrating the principles of fine-tuning and dataset creation for specific NLP tasks.
• Rewriting AI-generated content: Can be used to make AI-produced text sound more natural and less robotic, particularly for passages between 80 and 300 words.

Limitations

• The model's output may lean towards a literary, slightly old-fashioned register due to its training data (pre-1929 prose).
• Best suited for English text and passages of 80 to 300 words; longer documents should be processed paragraph by paragraph.
• It is an educational demo and makes no claims regarding AI detector evasion.