In this assignment, I used the preference dataset created in LIxy839/DPO_preference_dataset, which contains 50 samples in the format {"prompt": "", "chosen": "", "reject": ""}. Based on this dataset, I fine-tuned the “mistralai/Mistral-7B-Instruct-v0.2” model using the Direct Preference Optimization (DPO) method. The training was run for 100 steps, with the model updated every 4 steps, ensuring gradual and stable improvement.

The file model_comparison_results.csv contains the results of 10 unseen instructions, along with the responses generated by both the original model and the DPO fine-tuned model, which has been pushed to Hugging Face Hub. Additionally, The resulting PEFT adapter has been pushed to Hugging Face Hub. The complete implementation of the DPO fine-tuning workflow, including all training steps, tokenizer and formatting processes, and model response comparisons, is provided in the uploaded Jupyter Notebook file dpo.ipynb.

Quick start

python

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="LIxy839/dpo-mistral-7b", 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 DPO, a method introduced in Direct Preference Optimization: Your Language Model is Secretly a Reward Model.

Framework versions

TRL: 0.17.0
Transformers: 4.51.3
Pytorch: 2.6.0
Datasets: 3.6.0
Tokenizers: 0.21.1

Citations

Cite DPO as:

bibtex

@inproceedings{rafailov2023direct,
    title        = {{Direct Preference Optimization: Your Language Model is Secretly a Reward Model}},
    author       = {Rafael Rafailov and Archit Sharma and Eric Mitchell and Christopher D. Manning and Stefano Ermon and Chelsea Finn},
    year         = 2023,
    booktitle    = {Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023},
    url          = {http://papers.nips.cc/paper_files/paper/2023/hash/a85b405ed65c6477a4fe8302b5e06ce7-Abstract-Conference.html},
    editor       = {Alice Oh and Tristan Naumann and Amir Globerson and Kate Saenko and Moritz Hardt and Sergey Levine},
}

Cite TRL as:

bibtex

@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 Gallou{\'e}dec},
	year         = 2020,
	journal      = {GitHub repository},
	publisher    = {GitHub},
	howpublished = {\url{https://github.com/huggingface/trl}}
}