novel-humanize-4b

novel-humanize-4b

Rewrites Chinese web-novel chapters so they read less like AI output, keeping plot, dialogue and characters as they are. A 4B model built on Qwen3-4B, for non-commercial use.

Made by gugujiji on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: novel-humanize-4b rewrites Chinese web-novel chapters whose prose reads as machine-made. The plot, the dialogue and the characters stay the same, and only the writing style changes. It is built on Qwen3-4B, and the card says an ordinary graphics card can run it.

Who it suits: It suits writers and editors who work on Chinese web fiction and want to smooth out stiff AI-style wording. The license is non-commercial, so it fits personal and research use.

Title: Who runs it
Industry | Who uses it | How they use it | Example
Publishing | Web-novel authors | Rework AI-assisted chapters into a plainer voice | Rewrite one chapter before posting it
Editing | Editors | Tidy flowery phrasing and stock metaphors | Pass a draft chapter through the model
Research | Text generation researchers | Study style rewriting with a small model | Compare outputs under the card's settings

If you work with other genres or other languages, you can skip it. The card does not promise results outside Chinese web novels.

Where to use it:

• A vLLM server on a machine with a GPU, which the card uses for serving
• Python scripts that call an OpenAI-compatible endpoint
• Windows, macOS or Linux with Python, for downloading the files

Title: Jobs it does
Use case | Who | What they get back
Rewriting a chapter | Authors | The same story in a more natural voice
Fixing typos after a rewrite | Editors | Only the wrong characters corrected, with no new rewrite
Calling it as a tool | Developers | Rewrite and typo-fix tools that another model can call

How to set it up:

• Python on Windows, macOS or Linux
• A graphics card that can serve a 4B model with vLLM
• The fix_tools.py file from this repository, which holds the fixed system prompt and helpers
• For the optional typo step, an OpenAI-compatible API and its key

1. Install the Hugging Face Hub tools.
2. Download the model files.
3. Serve the model with vLLM, then send chapters with the card's fixed prompt.

pip install -U huggingface_hub
hf download gugujiji/novel-humanize-4b

vllm serve gugujiji/novel-humanize-4b --served-model-name novel-humanize-4b \
  --max-model-len 8192 --generation-config vllm

How to use it: Send one whole chapter at a time, about 1000 to 2500 characters. Use the card's system prompt and user message as written, with half-width punctuation, or import SYSTEM and wrap_input from fix_tools.py. For sampling, the card gives temperature 1.0 and top_p 0.95, adds top_k 20 and a presence penalty of 0.3, and turns thinking off.

Title: Parts the card names
Tool | What it does
SYSTEM | The fixed system prompt that tells the model how to rewrite
wrap_input | Wraps your chapter text in the fixed user message
rewrite_novel | A tool definition that rewrites a chapter
fix_typos | A tool definition that finds and fixes wrong characters only

from fix_tools import TOOLS, call_tool

msgs = [{"role": "user", "content": "帮我把这章去一下 AI 味：……"}]
r = client.chat.completions.create(model="deepseek-chat", messages=msgs, tools=TOOLS)
m = r.choices[0].message
if m.tool_calls:
    msgs.append(m)
    for tc in m.tool_calls:
        msgs.append({"role": "tool", "tool_call_id": tc.id,
                     "content": call_tool(tc.function.name, tc.function.arguments)})
    r = client.chat.completions.create(model="deepseek-chat", messages=msgs, tools=TOOLS)

Example: You paste a 1,500-character chapter full of ornate wording and stock similes. The model returns the same chapter with plainer phrasing and about the same length. The plot, the dialogue and the order of events stay put, and it stops where your text stops.

For newcomers and experts: If you are new, think of this as an editing pass that swaps showy wording for a plainer author's voice. The first small thing to try is one short chapter, after you have the model served.

If you know the field, the model is a Qwen3-4B fine-tune that must see the card's exact prompt. A rewrite can leave typos or mix up the particles of, 地 and 得, so fix_tools.py can run a second pass through your own model that corrects only the wrong characters. The trade-off is fidelity: it sometimes drops small details.

Serve the model with vLLM
Wrap one chapter in the fixed prompt
Generate with thinking off
Optionally fix typos with your own API

Good to know:

• Length: a chapter of 1000 to 2500 characters works best. Extremely short text invites invented plot, and long text should be split by paragraph.
• Details: it sometimes removes small details, so read the result before use.
• Genre: it targets Chinese web novels only, and other styles are not promised.
• License: cc-by-nc-4.0, with attribution required and commercial use banned, including paid services and commercial APIs.
• Cost to run: the rewrite needs a GPU. The optional typo step calls an OpenAI-compatible API with your own key, and the card's example uses deepseek-chat.
• Data it reaches: chapter text goes to your local server, and to your chosen API if you use the typo step.
• Not for: commercial use, which needs the author's permission first.

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 4B
• Weights: about 8 GB in BF16; about 2 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,690
• Likes: 86
• License: cc-by-nc-4.0
• Languages: zh
• Based on: Qwen/Qwen3-4B
• Quantized versions: 9
• Tags: safetensors, qwen3, chinese, novel, rewriting, text-generation, conversational

Links

• Model: https://huggingface.co/gugujiji/novel-humanize-4b
• Files: https://huggingface.co/gugujiji/novel-humanize-4b/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download gugujiji/novel-humanize-4b

Then follow the usage in its model card. Nothing in this kit runs by itself.