Sulthan Abiyyu Hakim, Rizal Setya Perdana, Tirana Noor Fatyanosa. Proceedings of the Second Workshop in South East Asian Language Processing, Association for Computational Linguistics, January 2025.
The gap
Indonesian LLMs are bad at distinguishing ethical instructions from unethical ones, and there was no Indonesian dataset to fix that. Alignment resources are overwhelmingly English, and what a model learns about harm in English does not transfer cleanly into another language and culture.
The dataset
Anak Baik means “good boy”. The name is the thesis: a well-behaved child refrains from harmful action, and that is the behavior we want a model to learn.
It is a set of Indonesian instruction-response pairs for supervised fine-tuning that deliberately includes both ethical and unethical responses. The point is to let a model learn the distinction, rather than memorize a list of refusal phrases. A model that has only ever seen refusals learns to refuse, not to reason.
Results
Fine-tuning Komodo and Cendol with LoRA produced significant improvement in ethical decision-making, validated with substantial gains in BLEU and ROUGE against socially responsible reference behavior.
The constraint that shaped the whole project was cost. Ethical alignment work for Indonesian cannot depend on a frontier-lab budget, or it will not happen at all. LoRA on existing open models, with a carefully curated dataset instead of a huge one, was the design that made it feasible.