Hugging Face Model Humanizer Rewrites AI Drafts
The Hugging Face model jialinyyzz/humanizer is a 12B parameter model that rewrites AI-written drafts to read like human-written text. It is designed to preserve every number, unit, date, name, and quote unchanged while making the prose sound natural.
How the Model Rewrites AI-Written Drafts
The model takes AI-generated text such as emails, essays, reports, or forum posts in English or Chinese and rewrites it to sound like a person wrote it. It follows a strict instruction to vary sentence length, cut hedging and throat-clearing, and prefer concrete words over abstract ones. The model does not add or remove any factual information, ensuring all numbers, dates, names, and quotes remain exactly as in the original draft. It runs locally and can be used offline, either through a downloadable app for Mac or Windows or via command-line tools like llama.cpp and AGENTS.md for coding agents.
Key Features and Benefits of the Model
The model has a 95% success rate in rewriting AI-written drafts to read like human-written text, as judged by Originality.ai at its strictest setting. In evaluations, 376 out of 420 English rewrites had no factual problems according to a strict LLM judge, and when issues were found, more than 9 in 10 were single-word or phrase fixes. The model is available in multiple GGUF file sizes, including Q8_0, Q6_K, Q4_K_M, Q3, and 2-bit versions, allowing users to choose based on their hardware, from high-end machines down to systems with as little as 8 GB of memory. All GGUF files are stored in the repository jialinyyzz/humanizer-GGUF and include built-in sampling settings for consistent performance.
Usage and Installation of the Model
Users can download the model and prompt format from Hugging Face using tools like hf download or brew and winget for llama.cpp. The app, available as a .dmg for Mac with Apple silicon or a Windows installer, suggests an appropriate model size on first run based on available memory and downloads it once for offline use. For command-line use with long documents or AI agents, users can install via pipx and run commands like hz paper.md -o paper.out.md. The model also supports integration with AI agents through AGENTS.md, which provides exact file paths, server commands, and prompt specifications. Sampling settings are fixed at temperature 1.0 and top-p 0.95, with top-k and min-p set to zero to ensure consistent behavior across platforms.
Evaluation and Results of the Model
According to evaluation data from the model’s held-out set, the strict fact judge flagged 44 of 420 English rewrites for factual issues, totaling 135 problems, 125 of which were single words or phrases. Originality.ai judged 95% of the rewrites as human-written, with only 11 of 210 samples flagged as AI at its strictest setting. The model’s quantization-aware training process, particularly for smaller files like the 2-bit and Q3 versions, improves token prediction accuracy from 70% to 87% compared to the base bf16 model. All GGUF files are standard builds compatible with recent versions of llama.cpp, and the model does not use any AI detector in its training process.
Technical Implementation and Prompt Format
The model uses a fixed prompt structure defined in prompt_format.json, where the instruction text is combined with the user’s draft and a separator to generate rewrites. The exact prompt format requires the instruction, followed by two newlines, the stripped draft, two newlines, and the string “### Rewritten:\n\n”. This structure must be reproduced byte-for-byte, as verified by the SHA-256 hash of the prompt for the draft “X” being cc51d66b4c593fbe. The model does not use system prompts, turn markers, or stop strings, and generates text until the end-of-sequence token is reached.
Sampling parameters are strictly controlled: temperature is set to 1.0, top-p to 0.95, and both top-k and min-p are explicitly set to zero to override default behaviors in backends like llama.cpp. These settings are stored in the metadata of all GGUF files released on or after 2026-10-04, ensuring consistent performance across applications without requiring manual configuration. The context window supports up to 8192 tokens for the combined instruction, draft, and output, with long documents processed by splitting at paragraph boundaries.
For integration with AI agents or custom workflows, the AGENTS.md file provides precise specifications, including file paths, server commands, and the exact prompt byte sequence. The model supports both completion and chat endpoints when using GGUF files with the built-in chat template, which reconstructs the required prompt from the last user message while ignoring system prompts and prior turns. This enables compatibility with tools like LM Studio’s Chat tab, provided they use the file’s template.
Model Availability and Repository Structure
The Hugging Face model jialinyyzz/humanizer is hosted alongside a dedicated GGUF repository at jialinyyzz/humanizer-GGUF, which contains all quantized versions of the model for local deployment. This separation allows users to access the base model architecture and its optimized GGUF files independently, ensuring compatibility with various inference backends like llama.cpp and MLX. The GGUF repository includes builds ranging from high-precision Q8_0 to ultra-low-bit 2-bit versions, each tailored for different hardware constraints.
All GGUF files in the jialinyyzz/humanizer-GGUF repository are standard builds compatible with recent versions of llama.cpp and include embedded sampling settings (temperature 1.0, top-p 0.95, top-k 0, min-p 0) in their metadata since 2026-10-04. These settings eliminate the need for manual configuration when using the model through applications or command-line tools. The repository also maintains sha256 checksums for each file, documented in USAGE.md, to verify integrity during download.
Users can access both repositories through standard Hugging Face tools: the base model via hf download jialinyyzz/humanizer and the GGUF files via jialinyyzz/humanizer-GGUF. This dual-repo setup supports workflows where developers pull the prompt format and model card from the main repository while retrieving optimized weights from the GGUF-specific repo, preserving existing download commands and ensuring backward compatibility.
Sources
Photo: unknown / rawpixel (CC CC0 1.0)
