Research · experiment
LoRA and QLoRA for Vietnamese News Summarization
A research note for comparing parameter-efficient fine-tuning strategies on Vietnamese summarization datasets.
Problem
Vietnamese news summarization needs strong factual compression while preserving names, numbers, and event relationships.
Research question
How far can parameter-efficient fine-tuning go before full fine-tuning becomes necessary?
Method
Compare LoRA and QLoRA runs across a consistent preprocessing pipeline, validation split, and evaluation protocol.
Evaluation
Start with ROUGE and BERTScore, then add manual checks for factual consistency and omitted critical details.
My observations
The interesting work is not only model choice. Dataset cleaning, summary length control, and evaluation design may dominate the final quality.