STUDI KOMPARATIF METODE PARAMETER-EFFICIENT FINE-TUNING UNTUK ABSTRACTIVE TEXT SUMMARIZATION

Muhammad Syafiq Akmal, NIM.: 24206052002 (2026) STUDI KOMPARATIF METODE PARAMETER-EFFICIENT FINE-TUNING UNTUK ABSTRACTIVE TEXT SUMMARIZATION. Masters thesis, UIN SUNAN KALIJAGA YOGYAKARTA.

Preview	Text (STUDI KOMPARATIF METODE PARAMETER-EFFICIENT FINE-TUNING UNTUK ABSTRACTIVE TEXT SUMMARIZATION) 24206052002_BAB-I_IV-atau-V_DAFTAR-PUSTAKA.pdf - Published Version Download (6MB) | Preview
	Text (STUDI KOMPARATIF METODE PARAMETER-EFFICIENT FINE-TUNING UNTUK ABSTRACTIVE TEXT SUMMARIZATION) 24206052002_BAB-II_sampai_SEBELUM-BAB-TERAKHIR.pdf - Published Version Restricted to Registered users only Download (10MB) | Request a copy

Abstract

Abstractive text summarization for Indonesian still faces high computational costs when adapting large pretrained language models through full fine-tuning, while prior studies on the Indonesian XL-Sum dataset have only tested full fine-tuning without comparing it to Parameter-Efficient Fine-Tuning (PEFT) methods such as LoRA, AdaLoRA, or QLoRA. This study aims to compare the summarization quality and computational efficiency of full fine-tuning against three PEFT methods, namely LoRA, AdaLoRA, and QLoRA, applied to the mT5-base backbone for Indonesian abstractive text summarization. Experiments were conducted on the Indonesian subset of the XL-Sum dataset, split into 38,242 training, 4,780 validation, and 4,780 test document-summary pairs; summarization quality was evaluated using ROUGE-1, ROUGE-2, ROUGE-L, BERTScore, and METEOR, supplemented with Wilcoxon signed-rank and paired bootstrap significance tests, while computational efficiency was measured through the trainable parameter ratio, peak GPU memory usage, training time, and inference time. The results show that all three PEFT methods consistently achieved better summarization quality than full fine-tuning across all measured metrics, with QLoRA reaching the highest scores on most metrics, while AdaLoRA was the most parameter- and training-time-efficient method and QLoRA was the most GPU-memory-efficient despite requiring the longest training time. These findings indicate that PEFT methods, particularly LoRA and QLoRA, can serve as more efficient and higher-quality alternatives to full fine-tuning for Indonesian text summarization, although the optimal method choice still depends on whether quality, memory efficiency, or training speed is prioritized, since no single method dominates across every dimension.

Share this knowledge with your friends :

Actions (login required)

View Item