<![CDATA[Informal Indonesian text poses challenges for NLP systems due to abbreviations, slang, and incomplete structures that deviate from standard language forms. This study evaluates the effectiveness of parameter-efficient fine-tuning (PEFT) using Low-Rank Adaptation (LoRA) to adapt large language models for informal-to-formal text normalization in a domain-specific context. Two open-weight multilingual models, Qwen3-14B and Mistral3-14B, were fine-tuned on a dataset of 1,233 formal sentences and 6.165 informal variations derived from academic information at Institut Teknologi Sepuluh Nopember (ITS). Evaluation on a held-out test set of 200 informal queries using lexical (BLEU, ROUGE) and semantic (cosine similarity and sentence-transformer-based similarity) metrics. The results show significant improvements after fine-tuning. Qwen3-14B achieved a BLEU score of 0.660, ROUGE-2 of 0.720, cosine similarity of 0.774, and semantic similarity of 0.907, while Mistral3-14B achieved 0.540, 0.624, 0.699, and 0.884, respectively. Additional experiments indicate that varying LoRA rank has minimal impact on performance, suggesting that LoRA configurations are sufficient for this task. These findings demonstrate that LoRA-based parameter-efficient fine-tuning enables effective and computationally efficient adaptation of large language models for Indonesian informal text normalization. This study provides a scalable solution for domain-specific NLP applications in low-resource settings, particularly for languages with high informal variation.]]>
