humanizer

humanizer

Rewrites AI-written drafts in English or Chinese so the text sounds human, keeping figures, dates, names and quotations as they are, with a way to lock chosen passages word for word.

Made by jialinyyzz on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: humanizer is a Gradio app that jialinyyzz published as a Hugging Face Space. It runs a 12B model that rewrites drafts an AI wrote, such as emails, essays, reports and forum posts, in English or Chinese, so the text sounds human. The model is trained to leave figures, dates, names and quotations untouched and to add nothing of its own.

Who it suits: People who start from an AI draft and want it to sound more natural without changing the facts in it. That covers writers, office staff and developers who work in English or Chinese.

Industry | Who uses it | How they use it | Example
Business | Office staff | Rewrite an AI-written email or report before it goes out | Softening a stiff status update while its dates stay put
Publishing | Forum and blog writers | Rewrite a post draft in English or Chinese | Turning a flat AI post into one that sounds natural
Software development | Developers | Send drafts to the /humanize API from a script | Looping through several drafts with the Python client

Skip it if you need text written from scratch or checked for truth, since it only rewrites a draft you bring.

Where to use it: humanizer works in these places.

• A web browser: the Space on Hugging Face runs the model on ZeroGPU
• A desktop app: macOS on Apple silicon and Windows, linked from the Space page
• A command line: the hz tool, plus steps for llama.cpp, MLX, transformers and vLLM
• Python scripts: the /humanize API through gradio_client

Use case | Who | What they get back
Rewrite a draft | Anyone with an AI-written text | A rewrite that keeps the facts, with a copy rate and a list of any numbers that went missing
Protect a passage | Writers with exact wording | A selected passage that comes through the rewrite word for word
Import a file | People with a .docx or PDF | The text of the file, up to 10 MB, loaded into the draft
Call the API | Developers | The rewrite from the /humanize endpoint

How to set it up: Online use needs only a browser. A local copy needs a few more things first.

• A web browser, for the online Space
• Python on Windows, macOS or Linux, for a local copy
• The Python packages listed in requirements.txt

1. Open the Space's page on Hugging Face if you only want to use it online.
2. For a local copy, install the Hugging Face hub library with pip.
3. Download the Space's files into a folder called humanizer.
4. Move into that folder.
5. Install Gradio at the version the Space uses.
6. Install the packages listed in requirements.txt.
7. Start the app with Python.

pip install -U huggingface_hub
hf download jialinyyzz/humanizer --repo-type space --local-dir humanizer
cd humanizer
pip install gradio==6.29.0
pip install -r requirements.txt
python app.py

How to use it: Online, open the Space's page, add your draft or import a .docx or PDF file, and run the rewrite. To protect a passage, select it and click Create fact first. With a local copy, start the app from its folder with the command below.

python app.py

Developers can call the rewrite from Python with gradio_client, using the call below.

Client("jialinyyzz/humanizer").predict(draft, api_name="/humanize")

Tool | What it does
Create fact | Keeps a selected passage word for word by swapping it for a placeholder that the model copies through
Import | Pulls the words out of a .docx or PDF file of up to 10 MB and discards the file straight after
/humanize API | Rewrites one part of up to 700 words per call and returns a copy rate and any missing numbers
python app.py | Starts the Gradio app from the main file

Example: You paste an AI-written email that holds a date and a person's name, select both and click Create fact. The page returns a more natural rewrite with those two passages unchanged. When a part drops a protected passage, the page redoes that part, up to 3 times, and never shows you the altered version.

For newcomers and experts: If you are new, paste one short email into the online Space and read the rewrite next to your original. Check the numbers, dates and names before you send anything.

If you know the field, the Space runs the model's bf16 weights on ZeroGPU. The cap for one run is roughly 2,100 words in English, or 3,600 characters in Chinese. Longer text is cut at paragraph breaks above 700 words or 1,200 characters, and each piece is rewritten in turn. If the GPU stops partway, the parts already done stay visible with a note. When the model answers an English draft in Chinese, the page samples that part again on its own.

You add a draft and mark passages to keep
Long text is cut at paragraph breaks
The model rewrites each part
A part that loses a kept passage is redone
You read the result

Good to know: It rewrites the text you give it, so the result still needs your own read.

• Accuracy: check the rewrite before sending it, mainly the numbers, dates and names, and whether each claim still points the same way
• Chinese: in the author's checks it trails English, with no factual problem in 149 of 204 rewrites
• Privacy: the Space stores submitted text and rewrites anonymously in a private dataset to improve the model, and does not share it with third parties
• Filter: emails, phone numbers, links, card numbers, IP addresses and ID numbers are replaced with placeholders by pattern matching, so names written out in words are not caught
• Removal: you can ask for your text to be deleted in the Space's Community tab, with the rough date and a short phrase from it
• Hardware: zero-a10g at Hugging Face
• Size: 20 files in the repository
• License: apache-2
• Cost to run: none named
• Data it reaches: the text you submit, stored as described above, and imported files, which are deleted once their text is read
• Not for: text you would not want stored, unless you run the model on your own computer

Title: English evaluation drafts flagged as AI, out of 210
Latest release: 11 drafts
Previous release: 26 drafts

On Hugging Face, 13 November 2026

• Likes: 51
• Built with: gradio
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/jialinyyzz/humanizer
• Files: https://huggingface.co/spaces/jialinyyzz/humanizer/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.