princeton-nlp / gemma-2-9b-it-DPO

gemma-2-9b-it-DPO Model Card

This model was trained under the same setup as gemma-2-9b-it-SimPO , with the DPO objective.

SimPO (Simple Preference Optimization) is an offline preference optimization algorithm designed to enhance the training of large language models (LLMs) with preference optimization datasets. SimPO aligns the reward function with the generation likelihood, eliminating the need for a reference model and incorporating a target reward margin to boost performance. Please refer to our preprint and github repo for more details.

Model Details

Model Description

We fine-tuned google/gemma-2-9b-it on princeton-nlp/gemma2-ultrafeedback-armorm with the DPO objective.

• Developed by: Yu Meng, Mengzhou Xia, Danqi Chen
• Model type: Causal Language Model
• License: gemma
• Finetuned from model: google/gemma-2-9b-it

Model Sources

• Repository: https://github.com/princeton-nlp/SimPO
• Paper: https://arxiv.org/pdf/2405.14734

How to Get Started with the Model

import torch
from transformers import pipeline

model_id = "princeton-nlp/gemma-2-9b-it-DPO"

generator = pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device="cuda",
)
outputs = generator([{"role": "user", "content": "What's the difference between llamas and alpacas?"}], do_sample=False, max_new_tokens=200)
print(outputs[0]['generated_text'])

Training Details

Training Data

We use princeton-nlp/gemma2-ultrafeedback-armorm as the preference optimization dataset.

Training Hyperparameters

We used the following hyperparameters:

• learning rate: 5e-7
• batch size: 128
• beta: 0.01

The other hyperparameters are kept the same with our SimPO recipe .

Speeds, Sizes, Times

Fine-tuning the google/gemma-2-9b-it on princeton-nlp/gemma2-ultrafeedback-armorm takes around 150 mins to finish on 8xH100 GPUs.

Evaluation Results

Technical Specifications

Model Architecture and Objective

The model architecture is based on google/gemma-2-9b-it . We use the DPO training objective.

Hardware

We used 8xH100 GPUs for model training.

Software

Training was done using the alignment-handbook library.

Citation

gemma model:

@article{gemma_2024,
    title={Gemma},
    url={https://www.kaggle.com/m/3301},
    DOI={10.34740/KAGGLE/M/3301},
    publisher={Kaggle},
    author={Gemma Team},
    year={2024}
}

DPO paper:

@article{rafailov2024direct,
  title={Direct Preference Optimization: Your language model is secretly a reward model},
  author={Rafailov, Rafael and Sharma, Archit and Mitchell, Eric and Manning, Christopher D and Ermon, Stefano and Finn, Chelsea},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  year={2024}
}

SimPO paper:

@article{meng2024simpo,
  title={{SimPO}: Simple preference optimization with a reference-free reward},
  author={Meng, Yu and Xia, Mengzhou and Chen, Danqi},
  journal={arXiv preprint arXiv:2405.14734},
  year={2024}
}

UltraFeedback paper:

@article{cui2023ultrafeedback,
  title={{UltraFeedback}: Boosting language models with high-quality feedback},
  author={Cui, Ganqu and Yuan, Lifan and Ding, Ning and Yao, Guanming and Zhu, Wei and Ni, Yuan and Xie, Guotong and Liu, Zhiyuan and Sun, Maosong},
  journal={arXiv preprint arXiv:2310.01377},
  year={2023}
}

ArmoRM paper:

@article{wang2024interpretable,
  title={Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts},
  author={Wang, Haoxiang and Xiong, Wei and Xie, Tengyang and Zhao, Han and Zhang, Tong},
  journal={arXiv preprint arXiv:2406.12845},
  year={2024}
}