PERBANDINGAN KUALITAS TEKNIK FINE TUNING LARGE LANGUAGE MODEL (LLM) DALAM TUGAS QUESTION ANSWERING

Miftah Firdaus, - and Yulia Retnowati, - and Mochamad Iqbal Ardimansyah, - (2025) PERBANDINGAN KUALITAS TEKNIK FINE TUNING LARGE LANGUAGE MODEL (LLM) DALAM TUGAS QUESTION ANSWERING. S1 thesis, Universitas Pendidikan Indonesia.

Abstract

Penelitian ini dilatarbelakangi oleh 2 permasalahan pada metode full fine-tuning, yaitu tingginya kebutuhan sumber daya komputasi dan bukti dari hasil studi sebelumnya bahwa perbedan teknik dapat mempengaruhi performa model pada tugas tertentu. Penelitian ini bertujuan memberikan rujukan akademis dalam pemilihan teknik parameter-efficient fine-tuning (PEFT) yang tepat untuk tugas Question Answering (QA) pada model BERT. Metode yang digunakan adalah eksperimen kuantitatif dengan membandingkan tiga teknik PEFT, yaitu LoRA, QLoRA, dan rsLoRA/RoRA dalam tugas QA menggunakan dataset SQuAD. Setiap metode dijalankan sebanyak lima kali dengan random seeds berbeda dan hyperparameter terkontrol untuk memastikan reproduksibilitas. Evaluasi kinerja menggunakan metrik Exact Match dan F1 Score, dengan analisis statistik. Hasil menunjukan bahwa LoRA memiliki waktu pelatihan paling singkat meskipun tidak signifikan dibandingkan rsLoRA/RoRA namun lebih baik signifikan terhadap QLoRA. QLoRA unggul dalam efisiensi memori, 3 kali lebih sedikit dibandingkan dua teknik lainnya. Pada metrik Exact Match, tidak terdapat perbedaan hasil yang signifikan, sedangkan pada metrik F1 Score, rsLoRA/RoRA unggul signifikan (71,93%) melampaui QLoRA (67,03%) dan LoRA (65,96%) sekitar 4-6%. Temuan ini mengindikasikan bahwa rsLoRA/RoRA memberikan hasil lebih baik sehingga dapat menjadi pilihan untuk tugas QA dengan mempertimbangkan efisiensi sumber daya dan kualitas keluaran. amun, temuan ini masih terbatas pada eksperimen menggunakan BERT-base dengan dataset SQuAD, sehingga penelitian lanjutan pada ukuran model, dataset, dan jenis tugas yang berbeda diperlukan untuk memvalidasi generalisasi hasil ini. ----- This research is driven by two main factors: the high computational and parameter requirements of fully fine-tuning large language models, and evidence from prior studies showing that variations in fine-tuning techniques can impact performance on specific tasks. The study aims to serve as an academic reference for choosing the most suitable Parameter-Efficient Fine-Tuning (PEFT) method for Question Answering (QA) tasks. A quantitative experimental approach was employed to compare three PEFT techniques (LoRA, QLoRA, and rsLoRA/RoRA) on QA tasks using the SQuAD dataset. Each method was run five times with different random seeds under controlled hyperparameters to ensure reproducibility. Performance was assessed using the Exact Match (EM) and F1 Score metrics, followed by statistical analyses. Results indicate that LoRA yielded the fastest training times, with no significant difference compared to rsLoRA/RoRA, but showing a significant improvement over QLoRA in this aspect. QLoRA demonstrated the best memory efficiency, requiring roughly one-third of the memory used by the other two methods. While EM scores showed no statistically significant differences among the techniques, the F1 Score results favored rsLoRA/RoRA (71.93%), outperforming QLoRA (67.03%) and LoRA (65.96%) by approximately 4–6%. These outcomes suggest that rsLoRA/RoRA offers the best overall performance for QA tasks when balancing resource efficiency with output quality. However, these findings are limited to the experiments with BERT-base on the SQuAD dataset, and further studies on other model sizes, datasets, and tasks are needed to validate the generalizability of the results.

	Text S_RPL_2000388_Title.pdf Download (2MB)
	Text S_RPL_2000388_Chapter1.pdf Download (502kB)
	Text S_RPL_2000388_Chapter2.pdf Restricted to Staf Perpustakaan Download (1MB)
	Text S_RPL_2000388_Chapter3.pdf Download (904kB)
	Text S_RPL_2000388_Chapter4.pdf Restricted to Staf Perpustakaan Download (935kB)
	Text S_RPL_2000388_Chapter5.pdf Download (773kB)
	Text S_RPL_2000388_Appendix.pdf Restricted to Staf Perpustakaan Download (1MB)

Item Type:	Thesis (S1)
Additional Information:	https://scholar.google.com/citations?view_op=new_profile&hl=en ID Sinta Dosen Pembimbing: Yulia Retnowati: 6852573 Mochamad Iqbal Ardimansyah: 6658552
Uncontrolled Keywords:	fine-tuning, peft, question-answering, lora, quantized lora, rslora, fine-tuning technique
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science Q Science > QA Mathematics > QA76 Computer software T Technology > T Technology (General)
Divisions:	UPI Kampus cibiru > S1 Rekayasa Perangkaat Lunak
Depositing User:	Miftah Firdaus
Date Deposited:	17 Sep 2025 01:51
Last Modified:	17 Sep 2025 01:51
URI:	http://repository.upi.edu/id/eprint/137072

Actions (login required)

View Item