DeepDive404-2/humanizer

humanizer: AI-to-Human Text Rewriter

humanizer is a 12 billion parameter model engineered to transform AI-written drafts into text that reads as if a human wrote it. Developed by DeepDive404-2, this model is fine-tuned to preserve all factual details (numbers, units, dates, names, quotes) while enhancing natural language, varying sentence structure, and eliminating AI-typical phrasing. It supports both English and Chinese and is designed for efficient local execution.

Key Capabilities

• Human-like Rewriting: Rewrites AI-generated content (emails, essays, reports, forum posts) to sound natural and human, significantly reducing detection by AI content detectors (achieved 95% human-judged rate by Originality.ai at its strictest setting).
• Factual Fidelity: Strictly maintains all factual information, including numbers, units, dates, names, and quotations, with 376 out of 420 English rewrites having no factual problems in evaluations.
• Multilingual Support: Proficiently rewrites text in both English and Chinese.
• Local Execution: Optimized for running locally on various devices, with GGUF files available in multiple quantizations (e.g., Q8_0, Q6_K, Q4_K_M, Q3, 2-bit) to suit different memory configurations.
• Quantization-Aware Training: Features advanced quantization techniques for 2-bit, Q3, and Q4_K_M files, providing higher agreement with the full bf16 model at smaller file sizes.

Good For

• Content Creators: Humanizing AI-generated articles, blog posts, and social media content.
• Professionals: Rewriting AI-drafted emails, reports, and presentations to ensure a personal and authentic tone.
• Students: Refining AI-assisted essays and academic papers to meet human writing standards.
• Developers: Integrating into applications for automated text humanization, with support for llama.cpp, MLX, transformers, and vLLM runtimes.
• Users with Privacy Concerns: Performing text transformations entirely offline, ensuring data privacy.