Perbandingan Performa Large Language Model Dengan Presisi 8bit vs BF16

This paper presents Legal-Case LLM, an open-source, fine-tuned language model tailored for Indonesian human-trafficking jurisprudence. General-purpose large language models exhibit high fluency but risk factual hallucination and limited jurisprudential fidelity when applied to legal texts. The objective is to develop a reproducible model that improves factual recall, legal terminology use, and jurisprudential alignment for Indonesian trafficking cases. Methods: We assembled a curated corpus of 400 court decisions from the Direktori Putusan Mahkamah Agung, extracted structured metadata and summaries, and generated question–answer pairs via large models followed by multi-stage cleaning and expert validation. We fine-tuned open models from the LLaMA family variants using parameter-efficient techniques (LoRA), evaluated with automatic metrics (ROUGE, BLEU, BERTScore, BARTScore), and a focused qualitative audit. Results: The fine-tuned model demonstrates marked improvements in content recall and semantic alignment versus zero-shot baselines, produces more jurisprudentially aligned phrasing (accurate use of terms such as amar putusan, Majelis Hakim, and percobaan), and reduces hallucination propensity in statute-related outputs. Conclusion and impact: Legal-Case LLM offers a reproducible, transparent tool to assist legal practitioners and researchers in Indonesia, while emphasising human-in-the-loop verification and citation-matching to ensure legal reliability and ethical deployment.