This repository contains a LoRA adapter trained for Humanization by Iterative Paraphrasing (HIP), from the paper Base Models Look Human To AI Detectors.

The adapter is intended to be loaded on top of:

Qwen/Qwen3-4B-Base

Model Details

Adapter type: LoRA / PEFT adapter
Base model: Qwen/Qwen3-4B-Base
Training objective: AI-to-human paraphrase reconstruction
Training data: paired AI-style and human-written passages from the HIP training data
Intended pipeline: iterative paraphrasing, where the adapter rewrites the previous round's output for a fixed number of rounds

Intended Use

This adapter is released to support research reproducibility for the HIP paper. It is intended for studying detector behavior, paraphrase-based rewriting, and robustness of AI-text detectors.

The adapter should not be used to evade deployed academic-integrity, authorship, or provenance systems in real-world settings.

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "Qwen/Qwen3-4B-Base"
adapter = "YixuanEvenXu/Qwen3-4B-Base-HIP-adapter"

tokenizer = AutoTokenizer.from_pretrained(adapter)
model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

For the full minimal HIP pipeline, see the code release linked from the paper.

Training Summary

The adapter was trained with supervised fine-tuning on paired examples (a_i, h_i), where a_i is an AI-style paraphrase of a human passage and h_i is the corresponding human-written target. Training uses a plain source-target format rather than a chat template.

Citation

If you use this adapter, please cite:

bibtex

@article{xu2026base,
  title={Base Models Look Human To AI Detectors},
  author={Yixuan Even Xu and Ziqian Zhong and Aditi Raghunathan and Fei Fang and J. Zico Kolter},
  journal={arXiv preprint arXiv:2605.19516},
  year={2026}
}

License and Terms

This adapter is released under Apache License 2.0, following the license of the Qwen3 base model family. Users are responsible for complying with the terms of the base model and any downstream deployment requirements.

Base model reference:

Qwen/Qwen3-4B-Base

Limitations

This adapter was trained for a specific research setting and evaluated on selected English prose domains. Performance may differ across domains, languages, detectors, and future detector versions.