relation: http://repository.upi.edu/141950/
title: PENDEKATAN BERKELANJUTAN BERBASIS PARAMETER-EFFICIENT FINE-TUNING UNTUK MEMBANGUN SISTEM PERCAKAPAN BERBASIS AI DENGAN DUKUNGAN BAHASA DAERAH PADA LARGE VISION LANGUAGE MODEL
creator: Ramadhirra Azzahra Putri, -
creator: Liptia Venica, -
creator: Dewi Indriati Hadi Putri, -
subject: T Technology (General)
description: Indonesia memiliki >700 bahasa daerah yang kurang terreprentasi secara digital, sehingga adaptasi Large Vision Language Models (LVLMs) perlu pendekatan yang kontekstual dan hemat sumber daya. Studi ini mengadaptasi dua backbone berbasis LLaMa 3.2 11B Vision dan Gemma 3 12B PT untuk bahasa Sunda dan Jawa melalui Parameter-Efficient Fine-Tuning (PEFT) skema QSBoRA-FA (matriks A dibekukan/di-one-hot, adaptasi pada B), assistant-only loss, kuantisasi 4-bit via Unsloth, serta pelacakan emisi dengan CodeCarbon. Rangkaian meliputi continued pretraining pada Cendol Collection, instruction tuning pada Alpaca Sundanese/Javanese (cleaned) dan korpus distillation (≈7k per bahasa), lalu Direct Preference Optimization (DPO) pada pasangan cho-sen/rejected (≈1k+1k). Pada NusaX-MT, LLaMa 3.2 11B Vision meraih chrF++ 60,50 (rata-rata 6 arah terjemahan: Sun↔Ind, Jav↔Ind, Sun↔Jav) dan Gemma 3 12B PT meraih skor 45,70; catatan bahwa pada skenario low-resource chrF++ ≥50 lazimnya sudah “sangat baik”, sehingga capaian 60,50 menandai kualitas terjemahan yang matang. Pada NusaX-Senti, LLaMa 3.2 11B Vision mencapai Weighted-F1-score 95,42% (akurasi 94,00%), sementara Gemma 3 12B PT men-capai Weighted-F1-score 89,76% (akurasi 95,00%); Weighted-F1-score ditekankan karena ketahanannya terhadap ketidakseimbangan label daripada akurasi yang cepat jenuh. Jejak operasional tercatat ≈21,854 kg CO₂e (energi 48,260 kWh, waktu 88,23 jam); tahap CP mendominasi konsumsi, sedangkan DPO berskala menit. Angka tersebut sekitar 0,096% dari emisi BLOOM-176B (24,7 ton) dan 0,0047% dari GPT-3 (502 ton), sehingga skema QSBoRA-FA (dengan Unsloth) menunjukkan adaptasi LVLM low-carbon tanpa trade-off kualitas yang berarti pada tugas target, sekaligus menawarkan baseline Green AI untuk konteks Indonesia.  -----  Indonesia hosts 700+ local languages that are digitally underrepresented, mak-ing Large Vision Language Model (LVLM) adaptation a problem of linguistic context and resource efficiency. We adapt two backbones, LLaMa 3.2 11B Vision and Gemma 3 12B PT, using QSBoRA-FA Parameter-Efficient Fine-Tuning (PEFT) (freezing/one-hotting matrix A, adapting B), assistant-only loss, 4-bit quantization via Unsloth, and CodeCarbon for emissions tracking. The pipeline comprises continued pretraining on Cendol Collection, instruction tuning on cleaned Alpaca Sundanese/Javanese and 7k/language distilled corpora, followed by Direct Preference Optimization (≈1k+1k chosen/rejected pairs). On NusaX-MT, LLaMa 3.2 11B Vision attains chrF++ 60.50 averaged over 6 translation di-rections (Sun↔Ind, Jav↔Ind, Sun↔Jav) while Gemma 3 12B PT scores 45.70; notably, in low-resource settings chrF++ ≥50 is commonly considered “very good,” so 60.50 indicates strong translation quality. On NusaX-Senti, LLaMa 3.2 11B Vision reaches 95.42% Weighted-F1-score (94.00% accuracy) and Gemma 3 12B PT 89.76% Weighted-F1-score (95.00% accuracy); we highlight Weighted-F1-score as IT is more robust to class imbalance than accuracy. The total opera-tional footprint is ≈21.854 kg CO₂e (48.260 kWh, 88.23 h), with CP dominating and DPO being minute-scale; this is ≈0.096% of BLOOM-176B (24.7 t) and ≈0.0047% of GPT-3 (~502 t). Overall, QSBoRA-FA (with Unsloth) delivers low-carbon LVLM adaptation without meaningful quality trade-offs on the target tasks, providing a practical Green AI baseline for the Indonesian context.
date: 2025-09-12
type: Thesis
type: NonPeerReviewed
format: text
language: id
identifier: http://repository.upi.edu/141950/8/S_MKB_2100188_Title.pdf
format: text
language: id
identifier: http://repository.upi.edu/141950/2/S_MKB_2100188_Chapter1.pdf
format: text
language: id
identifier: http://repository.upi.edu/141950/3/S_MKB_2100188_Chapter2.pdf
format: text
language: id
identifier: http://repository.upi.edu/141950/4/S_MKB_2100188_Chapter3.pdf
format: text
language: id
identifier: http://repository.upi.edu/141950/5/S_MKB_2100188_Chapter4.pdf
format: text
language: id
identifier: http://repository.upi.edu/141950/6/S_MKB_2100188_Chapter5.pdf
format: text
language: id
identifier: http://repository.upi.edu/141950/9/S_MKB_2100188_Appendix.pdf
identifier:   Ramadhirra Azzahra Putri, - and Liptia Venica, - and Dewi Indriati Hadi Putri, -  (2025) PENDEKATAN BERKELANJUTAN BERBASIS PARAMETER-EFFICIENT FINE-TUNING UNTUK MEMBANGUN SISTEM PERCAKAPAN BERBASIS AI DENGAN DUKUNGAN BAHASA DAERAH PADA LARGE VISION LANGUAGE MODEL.  S1 thesis, Universitas Pendidikan Indonesia.   
relation: https://repository.upi.edu/