vectorzhou/gemma-2-2b-it-alpaca-cleaned-SFT-PKU-SafeRLHF-OnlineIPO1-0317153039-epoch-3

Model Card for gemma-2-2b-it-alpaca-cleaned-SFT-PKU-SafeRLHF-OnlineIPO1

This model is a fine-tuned version of vectorzhou/gemma-2-2b-it-alpaca-cleaned-SFT on the PKU-Alignment/PKU-SafeRLHF dataset. It has been trained using TRL.

Quick start

from transformers import pipeline

question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
generator = pipeline("text-generation", model="vectorzhou/gemma-2-2b-it-alpaca-cleaned-SFT-PKU-SafeRLHF-OnlineIPO1-0317153039-epoch-3", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])

Training procedure

This model was trained with Extragradient, a method introduced in Extragradient Preference Optimization (EGPO): Beyond Last-Iterate Convergence for Nash Learning from Human Feedback.

Framework versions

• TRL: 0.13.0
• Transformers: 4.48.0
• Pytorch: 2.2.1
• Datasets: 3.2.0
• Tokenizers: 0.21.0

Citations

Cite Extragradient as:

@misc{zhou2025extragradientpreferenceoptimizationegpo,
    title={Extragradient Preference Optimization (EGPO): Beyond Last-Iterate Convergence for Nash Learning from Human Feedback}, 
    author={Runlong Zhou and Maryam Fazel and Simon S. Du},
    year={2025},
    eprint={2503.08942},
    archivePrefix={arXiv},
    primaryClass={cs.LG},
    url={https://arxiv.org/abs/2503.08942}, 
}

Cite TRL as:

@misc{vonwerra2022trl,
    title        = {{TRL: Transformer Reinforcement Learning}},
    author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallouédec},
    year         = 2020,
    journal      = {GitHub repository},
    publisher    = {GitHub},
    howpublished = {\url{https://github.com/huggingface/trl}}
}