This repository hosts a lightweight text summarization model fine-tuned from DistilGPT2 using Direct Preference Optimization (DPO). The model was trained on preference-labeled data to generate summaries that align better with human preferences compared to traditional supervised fine-tuning.

Model Details

Base model: DistilGPT2
Fine-tuning method: Direct Preference Optimization (DPO)
Dataset: Preference pairs with prompt, chosen, and rejected summaries
Evaluation metrics: ROUGE-1 (0.2841), ROUGE-L (0.2247), BLEU (0.0286)
Use case: Generating high-quality, human-aligned text summaries

How to Use

You can load and use the model easily with the Hugging Face Transformers library:

python

from transformers import pipeline

summarizer = pipeline("text-generation", model="justthzz/preference-tuned-summarizer")
text = "Summarize: Your input text here."

summary = summarizer(text, max_length=150, do_sample=False)
print(summary[0]['generated_text'])

Files Included

pytorch_model.bin - Model weights
config.json - Model configuration
Tokenizer files (tokenizer.json, vocab.txt, etc.)
Model card and this README

About DPO

Direct Preference Optimization is a fine-tuning technique that leverages preference-labeled datasets to directly optimize a model’s output preferences. This method improves alignment with human judgments beyond typical supervised fine-tuning.

Evaluation Results

Metric	Base Summary (avg)	DPO Summary (avg)
ROUGE-1	0.0442	0.2841
ROUGE-L	0.0366	0.2247
BLEU	0.0000	0.0286

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}}
}

justthzz/
preference-tuned-summarizer

Preference-Tuned Summarizer using Direct Preference Optimization (DPO)