Merak-7B-v2: An Indonesian Language LLM

Merak-7B-v2 is a 7 billion parameter large language model developed by Muhammad Ichsan, specifically designed for the Indonesian language. It is built upon the Meta Llama-2-7B-Chat-HF architecture and has been fine-tuned using QLoRA for efficient performance, enabling it to run with 16 GB VRAM.

Key Capabilities

• Indonesian Language Proficiency: The model is extensively fine-tuned on a dataset of 600,000 Indonesian Wikipedia articles, enhancing its understanding and generation capabilities in Bahasa Indonesia.
• Efficient Deployment: Utilizes QLoRA (Quantized LoRA) for efficient fine-tuning and inference, making it accessible on systems with limited VRAM (e.g., 16 GB).
• Prompt-Style Adaptation: The v2 iteration incorporates changes in prompt-style during fine-tuning, building upon the initial Merak-7B model which was trained on 200,000 Indonesian Wikipedia articles.

Good For

• Applications requiring strong performance in the Indonesian language.
• Researchers and developers working on Indonesian NLP tasks.
• Environments with VRAM constraints that benefit from 4-bit quantization, though higher VRAM is recommended for optimal answer quality.