Downloads · 30 days
249
22% of all-time downloads
gohumanize-open-humanizer-qlora
gohumanize/gohumanize-open-humanizer-qlora
gohumanize-open-humanizer-qlora is a text generation model from gohumanize. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as apache-2.0.
The version 1 QLoRA build of the GoHumanize Open Humanizer, a small open model that rewrites AI-styled English prose into more natural human writing. It was trained on the version 1 dataset (2,000 book passages) with…
All-time downloads
1.1K
Public
Parameters
4B
14.9 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors8.1 GB · 54%
At a glance
- Task
- Text Generation
- Library
- transformers
- License
- apache-2.0
- Model type
- qwen3
- Access
- Public
- Created
- Sep 22, 2026
- Updated
- Sep 23, 2026
- SHA
- e22990b2
Try a prompt
Base models
Model card
From the Hugging Face model README
GoHumanize Open Humanizer (QLoRA version)
The version 1 QLoRA build of the GoHumanize Open Humanizer, a small open model that rewrites AI-styled English prose into more natural human writing. It was trained on the version 1 dataset (2,000 book passages) with a small LoRA adapter, for anyone who wants the adapter or wants to reproduce training on a 24 GB GPU.
For most uses, take the main model instead. Version 2 of the main model adds modern prose to the training data and returns modern text nearly unchanged far less often (17% of outputs against 20% for this build and 38% for the version 1 full fine-tune; write-up, section 6.8).
Educational release, separate from the production systems of GoHumanize.ai. It makes no claim about AI detectors. Project page: https://gohumanize.ai/open-model
Files
- Root: merged 16-bit weights (the adapter already applied to Qwen3-4B), usable like any
transformersmodel. lora/: the LoRA adapter on its own (about 66 MB), to apply toQwen/Qwen3-4Bwith PEFT.gguf/: Q4_K_M and Q8_0 builds for Ollama, LM Studio and llama.cpp (ollama run hf.co/gohumanize/gohumanize-open-humanizer-qlora:Q4_K_M).
Usage, system prompt and chat format are the same as for the main model.
Training
| Base model | Qwen/Qwen3-4B (Apache-2.0) |
| Method | QLoRA with Unsloth: 4-bit base, LoRA rank 16, alpha 32, all attention and MLP projections |
| Data | 2,000 train / 200 test pairs, see the dataset card |
| Loss | on the assistant (human target) tokens only |
| Epochs, LR, batch | 2 epochs (250 steps), 2e-4 cosine, effective batch 16, max 1,024 tokens |
| Hardware | 1x NVIDIA A10G (24 GB) on Modal, 21.7 minutes |
| Tracking | Weights & Biases run |
| Final losses | train 1.30, eval 1.39 (3.17 before training) |
Evaluation
| Measure (200 held-out pairs) | Base Qwen3-4B | QLoRA (this repo) | Full fine-tune (main model) | Human |
|---|---|---|---|---|
| BERTScore F1 vs human | 0.900 | 0.921 | 0.920 | |
| ROUGE-L vs human | 0.424 | 0.540 | 0.535 | |
| Names kept (recall) | 0.594 | 0.623 | 0.614 | |
| Length ratio vs human | 0.83 | 0.93 | 0.95 | 1.00 |
| Stock LLM phrases / text | 0.01 | 0.00 | 0.00 | 0.00 |
| Average sentence length | 15.2 | 22.5 | 22.6 | 28.1 |
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 0.1.0. Zenodo. https://doi.org/10.5281/zenodo.22843083
GoHumanize, the product this research comes from: https://gohumanize.ai/
- Task
- Text Generation
- Library
- transformers
- Type
- qwen3
- License
- apache-2.0
- Languages
- en
- Created
- Sep 22, 2026
- Updated
- Sep 23, 2026