技术规格

Sahabat-AI-Llama3-8B-Instruct

Sahabat-AI-Llama3-8B-Instruct is a specialized large language model developed through a collaboration between GoTo Group and Indosat Ooredoo Hutchison. This model is constructed using a continued pre-training (CPT) approach on the Meta Llama 3 architecture, specifically optimized to reflect the linguistic patterns and cultural context of Indonesia. By incorporating a significant corpus of Indonesian text and regional languages such as Javanese and Sundanese, the model provides localized language processing capabilities that account for regional idioms and social contexts.

The technical framework is a dense, decoder-only Transformer architecture comprising 32 layers and a hidden dimension of 4096. It employs Grouped Query Attention (GQA) with 32 query heads and 8 key-value heads to improve inference efficiency. The model utilizes Rotary Positional Embeddings (RoPE) for sequence modeling and SwiGLU activation functions within its feed-forward layers. Training was facilitated by the NVIDIA NeMo framework, allowing the weights to be refined on a dataset of approximately 50 billion tokens, followed by supervised fine-tuning on hundreds of thousands of instruction-completion pairs.

This instruction-tuned variant is designed for high-quality interactions in both formal and informal Indonesian. It addresses specific cultural sensitivities and linguistic variations that are often missing in general-purpose global models. Primary applications include automated customer support for the Indonesian market, localized content synthesis, and technical assistance within the regional digital ecosystem. The model is compatible with the Transformers library and optimized for deployment on standardized accelerated computing infrastructure.

没有可用的 Sahabat-AI-Llama3-8B-Instruct 评估基准。