Pengembangan Large Language Model untuk Analisis Tindak Pidana Perdagangan Orang di Indonesia

The development of Large Language Models (LLMs) has demonstrated significant capabilities in understanding and generating text, but their application in the legal domain still faces serious challenges such as fact hallucination, linguistic bias, and low suitability for local jurisprudence structures. This research aims to develop a Legal-Case LLM, an open-weight large language model that is specifically fine-tuned for the analysis of jurisprudence on human trafficking in Indonesia. The dataset was compiled from more than 400 court decisions obtained from the Supreme Court Decision Directory through a process of scraping, document validation, structured metadata extraction, and the construction of synthetic question-and-answer (Q&A) data that was cleaned and validated in stages. The base model based on a decoder-only architecture (Gemma 3) was adapted using parameter-efficient fine-tuning (LoRA) techniques to improve training efficiency and domain suitability. The evaluation was conducted using automated metrics including ROUGE, BLEU, BERTScore, and BARTScore, as well as qualitative analysis of legal terminology alignment and hallucination tendencies. The results show that the fine-tuned model experienced a significant improvement in content recall, semantic alignment, and the use of legal terminology such as “court decision,” “panel of judges,” and “legal considerations” compared to the zero-shot baseline. In addition, the model was able to reduce factual errors in related outputs.