DahonoLabs/Dahono-4B

Dahono-4B: Sovereign Indonesian Legal AI

Dahono-4B, developed by DahonoLabs, is a 4.5 billion parameter model built on the Qwen3.5-4B-Base architecture, specifically tailored for the Indonesian legal domain. This v0.8 release significantly enhances performance and functionality, achieving an 89.40% Composite Legal Score and 89.80% Doctrinal Accuracy on the LegalBench-ID suite.

Key Capabilities & Innovations

• Specialized Legal Reasoning: Optimized for Indonesian law, compliance, and juridical analysis, with a focus on civil and labor law cases.
• Multimodal Document Vision: Integrated capability to process visual inputs (photos/scans) of legal documents to validate agreement conditions (e.g., Article 1320 Civil Code), signature authenticity, and stamp duty validity (Law No. 10/2020).
• Reasoning Reinforcement Learning (GRPO): Utilizes Group Relative Policy Optimization with four juridical reward functions to ensure format compliance, language purity, citation precision, and reasoning depth.
• Enhanced Language Purity: Achieves 100% purity in standard Indonesian legal language within its internal reasoning blocks, eliminating English code-switching.
• Robust Training Corpus: Trained on an expanded corpus of 6,271 samples, including 4,500 Chain-of-Thought reasoning samples and 1,771 legal glossary terms.

When to Use Dahono-4B

• Legal Research & Review: Accelerate research and review of Indonesian positive law documents.
• Compliance Pre-Checks: Conduct initial checks for regulatory compliance.
• Drafting Legal Instruments: Assist in drafting initial legal documents such as summons, cooperation agreements, and formal advocate correspondence.
• Legal Education: Serve as an educational tool for academics and legal practitioners.
• Local Inference: Available in GGUF Q4_K_M format for efficient local execution on devices with >= 4 GB RAM/VRAM via Ollama or llama.cpp.