Dahono-4B: A Sovereign Foundation Model for Indonesian Legal Reasoning, Statutory Hierarchy Alignment, and Multimodal Document Intelligence
Abstract
General-purpose Large Language Models (LLMs) predominantly trained on Anglo-American Common Law corpora demonstrate severe jurisdictional fragility when confronted with civil law traditions, frequently exhibiting statutory hallucination, citation of repealed legislation, and systematic disregard for statutory hierarchies (e.g., Indonesian Law No. 12/2011). In this paper, we introduce Dahono-4B (v0.8 Foundation Edition), a sovereign 4.02-billion parameter foundation language model specifically developed for Indonesian jurisprudence and multimodal document intelligence. Built upon an optimized Qwen-3.5 architecture, Dahono-4B integrates a dual-stage alignment regimen: supervised fine-tuning over structured Indonesian legal reasoning chains and Rule-Based Juridical Reinforcement (GRPO) governed by formal statutory citation verifiers. To empower empirical validation, we construct and release three foundational artifacts: (i) Indonesian-Legal-Reasoning-QA (5,000 multi-turn Chain-of-Thought pairs), (ii) Indonesian-Legal-Glossary (1,771 verified statutory concepts), and (iii) Indonesian-Legal-Vision (1,000 multimodal statutory gazette and privacy-compliant synthetic contract document images). On the standardized LegalBench-ID benchmark, Dahono-4B attains an overall Composite Legal Score of 88.82% (+14.32% over the unaligned base model), achieving 94.00% Hallucination Resistance, 89.20% Citation Precision, and 85.00% zero-shot visual document comprehension. Fully quantized to 2.49 GB (GGUF Q4_K_M), Dahono-4B guarantees absolute on-premise execution, securing client-attorney confidentiality under Indonesian Law No. 27/2022 (UU PDP).
