avasil02/automated-humanizer-lora

Automated Humanizer — LoRA adapter

LoRA adapter for Qwen/Qwen2.5-3B-Instruct that rewrites LLM-drafted academic text to read as human-authored, while preserving the scientific content. It is the default local_lora backend of the automated_humanizer pipeline, which pairs each rewrite with an AI-detector gate and an LLM academic-validity judge.

• —Base model: Qwen/Qwen2.5-3B-Instruct (load the base separately; this repo contains only the adapter)
• —Training data: synthetic (AI-draft, human-original) paragraph pairs built from award-winning pre-2020 arXiv papers (see the project's training/ subsystem)
• —Hardware: trained and runs fully offline on an NVIDIA Jetson Orin (~6 GB GPU/unified RAM at inference)

Usage

Intended to be used through the pipeline — see the project README. Standalone:

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-3B-Instruct")
model = PeftModel.from_pretrained(base, "avasil02/automated-humanizer-lora")
tokenizer = AutoTokenizer.from_pretrained("avasil02/automated-humanizer-lora")

Intended use & limitations

Edits style only; the surrounding pipeline verifies that the original results and claims are preserved. It is for making genuine research read naturally — not for misrepresenting authorship of the underlying work. Trained on English academic prose; quality outside that domain is untested.

Framework versions

• —PEFT 0.19.1