Qwen3-PUPPET-8B

<p align="center"> <img src="readme_assets/puppet_greeting.png" alt="PUPPET" width="200"> </p>

---

Qwen3-PUPPET-8B

Overview

This model is a fine-tuned version of Qwen/Qwen3-8B trained using PUPPET (GitHub) - a framework that jointly optimizes LLM output detectability and task performance via DPO.

This model is a research artifact released to accompany the paper, "LLM Output Detectability and Task Performance Can be Jointly Optimized" (Saito et al., arXiv, 2026).

Training Details

• Base Model: Qwen/Qwen3-8B [Apache 2.0]
• Training Data: Hello-SimpleAI/HC3 / reddit_eli5 split [CC BY-SA 4.0]
• Training Method: DPO (Direct Preference Optimization)
• Preference Labeling: Assign the best sample "chosen" and the worst sample "rejected" based on the sum of these scores following -
  • Task Performance - ROUGE-L
  • Detectability - openai-community/roberta-large-openai-detector [MIT]
• Training Environment: 2× NVIDIA RTX A6000 (48 GB VRAM each), TRL 0.24.0, Transformers 4.57.1

Warnings & Disclaimer

• This model is released for research purposes only.
• No guarantee is made regarding the accuracy or appropriateness of model outputs.
• Users are responsible for ensuring appropriate use in compliance with applicable laws.

License

This model is based on Qwen3-8B [Apache 2.0] by Qwen Team / Alibaba Cloud. This fine-tuned model is released under the Apache 2.0 License.

The training data (Hello-SimpleAI/HC3) is licensed under CC BY-SA 4.0. See: https://huggingface.co/datasets/Hello-SimpleAI/HC3

How to cite

If you find our code or work helpful, please cite:

@misc{Saito:PUPPET:2026,
  author        = {Koshiro Saito and Ryuto Koike and Masahiro Kaneko and Naoaki Okazaki},
  title         = {{LLM} Output Detectability and Task Performance Can be Jointly Optimized},
  eprint        = {2605.01350},
  howpublished  = {arXiv:2605.01350},
  primaryClass  = {cs.CL},
  year          = {2026},
}

---

Citation

@misc{qwen3,
      title={Qwen3 Technical Report}, 
      author={QwenTeam},
      year={2025},
      eprint={2505.09388},
      primaryClass={cs.CL},
      howpublished={arXiv:2505.09388}, 
}

@misc{hc3,
    title = "How Close is {C}hat{GPT} to Human Experts? Comparison Corpus, Evaluation, and Detection",
    author = "Biyang Guo and Xin Zhang and Ziyuan Wang and Minqi Jiang and Jinran Nie and Yuxuan Ding and Jianwei Yue and Yupeng Wu",
    primaryClass = {cs.CL},
    eprint = {2301.07597},
    howpublished = {arXiv:2301.07597},
    year = "2023",
}

@misc{openai_detector,
    title={Release strategies and the social impacts of language models},
    author={Irene Solaiman and Miles Brundage and Jack Clark and Amanda Askell and Ariel Herbert-Voss and Jeff Wu and Alec Radford and Gretchen Krueger and Jong Wook Kim and Sarah Kreps and others},
    primaryClass={cs.CL},
    howpublished={arXiv:1908.09203},
    year={2019},
    eprint={1908.09203},
}