Gemma2 9B CPT Sahabat-AI v1 Instruct Overview

Gemma2 9B CPT Sahabat-AI v1 Instruct is a 9 billion parameter instruction-tuned model developed by PT GoTo Gojek Tokopedia Tbk and AI Singapore, built upon the Gemma2 architecture. It is part of the Sahabat-AI ecosystem, a collection of LLMs focused on the Indonesian language and its various dialects. The model was fine-tuned using approximately 448,000 Indonesian instruction-completion pairs, 96,000 Javanese pairs, 98,000 Sundanese pairs, and an additional 129,000 English instruction-completion pairs, giving it strong multilingual capabilities.

Key Capabilities

• Multilingual Proficiency: Excels in Indonesian, Javanese, Sundanese, and English, demonstrating strong performance on tasks in these languages.
• Instruction Following: Evaluated with the IFEval dataset, showing robust instruction-following capabilities, particularly in Bahasa Indonesia.
• Benchmark Performance: Achieves leading scores on the SEA HELM (BHASA) evaluation benchmark for general language tasks across Indonesian, Javanese, and Sundanese, and tops the IndoMMLU benchmark for Indonesian language understanding. It also shows competitive performance on English tasks from the HuggingFace LLM Leaderboard.

Use Cases

This model is ideal for applications requiring nuanced understanding and generation in Indonesian and its regional dialects, as well as English. It is particularly suited for tasks such as question answering, sentiment analysis, summarization, and general instruction-following in these languages. Developers should note that the model has not been aligned for safety and requires further safety fine-tuning for production use.