humanizer-qwen3.6-27b-lora

humanizer-qwen3.6-27b-lora

LoRA adapter of unsloth/Qwen3.6-27B fine-tuned to rewrite AI-generated text so it reads like natural human writing — same meaning, human voice. QLoRA (4-bit, r=16, alpha=32) trained with Unsloth + TRL's SFTTrainer on ~5k paired (ChatGPT answer -> human answer) examples from Hello-SimpleAI/HC3, with a synthesized <think> trace so the model learns to spot machine-like patterns and plan the rewrite before answering.

Single-purpose model: it only humanizes existing text. Not a chat assistant.

Usage

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "DeFactOfficial/humanizer-qwen3.6-27b-lora",  # pulled from the Hub (adapter + 4-bit base) automatically
    max_seq_length=4096,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

Prompt with a system message asking for a humanizing rewrite and the passage in the user turn; the model thinks (<think>...</think>), then emits only the rewritten passage. See humanize.py in the training repo for the exact prompt and a chunked end-to-end script.

Training

• learning rate 0.0002 (cosine), effective batch 16, 1.0 epoch(s), max seq length 2048, adamw_8bit
• loss masked to the assistant turn only (train_on_responses_only)