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 transformers model.
  • lora/: the LoRA adapter on its own (about 66 MB), to apply to Qwen/Qwen3-4B with 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/

Downloads last month
240
Safetensors
Model size
4B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for gohumanize/gohumanize-open-humanizer-qlora

Finetuned
Qwen/Qwen3-4B
Adapter
(1,273)
this model

Dataset used to train gohumanize/gohumanize-open-humanizer-qlora