Optimizing Fine-Tuning Memory Usage with QLoRA for Automated Pedagogical Feedback Generation in Bahasa Indonesia

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Background: While recent advances in open-source large language models (LLMs) have enabled automated feedback generation for educational assessment, most other research is focused on English and requires high-end Graphics Processing Unit (GPU) resources. However, there is limited empirical evidence of the efficiency of LLMs in Bahasa Indonesia feedback tasks in low-resource settings like Indonesia, given the strict hardware limitations and differences in data sources Objective: This study investigates whether 4-bit quantization combined with Low-Rank Adaptation (LoRA)-based task adaptation can enable open-weight LLMs to generate Indonesian pedagogical feedback under strict 8GB Video Random Access Memory (VRAM) constraints, comparing the performance across various model architectures and supervision data (human-curated vs synthetic). Methods: A quantitative and experimental design using a 2 × 2 factorial structure was used to compare Llama-3-8B-Instruct and Gemma-2-9B-Instruct on 2 datasets of equal size (1,950 human-curated and 1,950 synthetic feedback pairs). For model fine-tuning, we implemented 4-bit Quantized Low-Rank Adaptation (QLoRA) quantization on consumer-grade GPUs with 8GB VRAM Performance was assessed by semantic similarity using Sentence-BERT (SBERT), similarity of the model with BLEU-4, ROUGE-L, peak VRAM usage, and training time. Results: QLoRA successfully enabled fine-tuning within 8GB VRAM, with peak memory usage between 7.0 and 7.8GB. Gemma-2-9B-Instruct achieved the highest semantic alignment (SBERT similarity = 0.77 on synthetic and 0.74 on human data), while Llama-3-8B-Instruct showed competitive performance (0.76 and 0.71, respectively). Human-curated datasets produced higher lexical alignment (ROUGE-1 up to 44.8) than synthetic datasets, though synthetic supervision remained semantically consistent. Conclusion: QLoRA enables efficient fine-tuning of large language models on consumer-grade hardware, making Indonesian feedback generation feasible in low-resource settings. While 4-bit quantization reduces memory requirements, LoRA enables task-specific pedagogical adaptation. Instruction-tuned models consistently outperform language-adapted base models. Future work should include human evaluation and broader educational contexts to improve generalizability. Keywords: Automated Feedback Generation, Large Language Models, QLoRA Fine-Tuning, Bahasa Indonesia, Educational Assessment