GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
gohumanize
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
gohumanize
Model Tree
Qwen/Qwen3-4B
Input Modalities
Output Modalities
Supported Functionality
A humanizer can fail by handing its input back almost unchanged. On 254 held-out passages, three tries each at temperature 0.9:
| Modern passages returned nearly unchanged (more than 90% of the words kept), from 30 articles never seen in training | 29.0% |
| Book passages returned nearly unchanged | 19.2% |
| Modern rewrites that lose a number | 22.1% |
| Invented datelines, footnote numbers or links, of 762 outputs | 0 |
The modern figures rest on 162 outputs from 54 passages, so a difference of a few points is within noise.
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.
Chat format, thinking disabled. The system prompt below is the one used in training.
python
from transformers import AutoTokenizer, AutoModelForCausalLMimport 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 three times in ten. 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/:
bash
ollama run hf.co/gohumanize/gohumanize-open-humanizer:Q4_K_MAlso 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.
| Base model | Qwen/Qwen3-4B (Apache-2.0) |
| Method | Full fine-tune with Unsloth: every weight updated in bf16, 8-bit AdamW, gradient checkpointing |
| Data | 2,629 train / 254 test pairs, see the dataset card |
| Loss | on the assistant (human target) tokens only |
| Epochs, LR, batch | 2 epochs, 2e-5 cosine, effective batch 16, max 1,024 tokens, seed 13 |
| Hardware | 1x NVIDIA H100 (80 GB) on Modal, 8 minutes of training |
| Tracking | Weights & Biases run |
| Final losses | train 0.99, eval 1.19 |
The untouched base model and the Open Humanizer on the 254 held-out passages (54 modern, 200 from books), against the human original.
| Measure (254 held-out pairs) | AI-styled input | Base Qwen3-4B | Open Humanizer | Human target |
|---|---|---|---|---|
| BERTScore F1 vs human (higher = closer meaning) | 0.923 | 0.905 | 0.931 | |
| ROUGE-L vs human (higher = closer wording) | 0.542 | 0.444 | 0.595 | |
| Names/capitalised tokens kept (recall) | 0.619 | 0.604 | 0.673 | |
| Length ratio vs human | 1.08 | 0.83 | 0.97 | 1.00 |
| Contractions per 100 words | 0.20 | 0.62 | 0.05 | 0.14 |
| Transition words per 100 words | 0.88 | 0.02 | 0.12 | 0.16 |
| Stock LLM phrases per text | 0.40 | 0.00 | 0.02 | 0.01 |
| Average sentence length (words) | 22.6 | 15.2 | 22.6 | 26.7 |
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.
| Resource | Link |
|---|---|
| Project page and browser demo | gohumanize.ai/open-model |
| GoHumanize (the product this research comes from) | gohumanize.ai |
| Model weights and GGUF builds | gohumanize/gohumanize-open-humanizer |
| Dataset, 2,883 pairs (CC-BY 4.0) | gohumanize/gohumanize-open-humanizer-dataset |
| Dataset on Kaggle | gohumanize/gohumanize-open-humanizer-dataset |
| Code and full pipeline | GoHumanize-ai/gohumanize-open-humanizer |
| Write-up: every step, service and result | docs/paper.md |
| Archived release, citable DOI | 10.5281/zenodo.22843083 |
| Python client and CLI | pypi.org/project/gohumanize-open-humanizer |
| MCP server for AI assistants | npm · source |
| Training run, loss curves and config | Weights & Biases |
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 0.4.0. Zenodo. https://doi.org/10.5281/zenodo.22843083
from transformers import AutoTokenizer, AutoModelForCausalLMimport 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))