gohumanize-open-humanizer

GoHumanize Open Humanizer

A small open model that rewrites AI-styled English prose into more natural human writing. It is a full fine-tune of Qwen3-4B (all four billion weights updated) on 2,957 pairs of (AI-styled passage, human original). The human side is public domain: 2,000 passages from books on Project Gutenberg and 957 passages of modern prose from US federal agencies.

The Open Humanizer is a public research and educational model created to demonstrate the general approach used to develop AI text humanization systems. It is separate from the production models used by GoHumanize.ai, but it reflects many of the same high-level principles: careful dataset preparation, transformation of source text into training pairs, model fine-tuning, evaluation and iterative improvement. It is not intended to reproduce the architecture, data, configuration or performance of the production systems, and it makes no claim about AI detectors.

Everything about the project (dataset, code, evaluation, write-up): https://gohumanize.ai/open-model

How often it rewrites

A humanizer can fail by handing its input back almost unchanged. On 300 held-out passages, three tries each at temperature 0.9:

The training pairs were built so the model learns to restructure sentences, not just swap words, and press-release datelines and footnote numbers were removed from the targets so the model does not learn to invent them. The write-up explains how.

Use

Chat format, thinking disabled. The system prompt below is the one used in training.

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

repo = "gohumanize/gohumanize-open-humanizer"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype=torch.bfloat16, device_map="auto")

SYSTEM = ("Rewrite the following text so that it reads as if a person wrote it: varied sentence "
          "length, concrete wording, natural rhythm, no filler transitions. Keep the meaning, the "
          "facts and the order of ideas. Return only the rewritten text.")
messages = [{"role": "system", "content": SYSTEM},
            {"role": "user", "content": "It is worth noting that the committee ultimately reached a consensus."}]
ids = tok.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=600, temperature=0.9, top_p=0.9, do_sample=True)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

On modern text, a single try still comes back nearly unchanged about one time in six. Generating a few tries and keeping the one that changed most (the MCP server and Python client do this for you) takes that close to zero.

Run it locally without a GPU using the GGUF builds in gguf/:

ollama run hf.co/gohumanize/gohumanize-open-humanizer:Q4_K_M

Also available: an MCP server (npx gohumanize-open-humanizer-mcp), a Python client (pip install gohumanize-open-humanizer) and a browser demo at https://gohumanize.ai/open-model.

Training

Evaluation

The untouched base model and the Open Humanizer on the 300 held-out passages (100 modern, 200 from books), against the human original.

These measures cannot see whether the model rewrote at all: an output identical to the AI-styled input scores well on meaning and wording. That is why the rewrite rate above is measured separately.

These measure how far the output moves from AI-styled prose towards the human target. They are not detector scores.

Limitations

• Two thirds of the human targets are pre-1929 books and the rest is US government writing, so the model leans literary on one side and official on the other. Casual modern writing is the gap.
• About three in ten rewrites of number-heavy text lose at least one figure. Some of that is rewording ("20 percent" as "a fifth"), some is loss. Check figures and other facts against the source.
• Trained on 80 to 300 word passages; rewrite long documents paragraph by paragraph.
• English only.
• Educational demo; no claims about AI detectors.

Links

Licence and citation

Apache-2.0. Cite as:

GoHumanize team (2026). GoHumanize Open Humanizer: an open text-humanization model, dataset and pipeline built from public-domain data. Version 0.2.0. Zenodo. https://doi.org/10.5281/zenodo.22843083