Model Overview

Sahabat-AI/llama3-8b-cpt-sahabatai-v1-instruct is an 8 billion parameter instruction-tuned model built on the Llama3 architecture, developed by PT GoTo Gojek Tokopedia Tbk and AI Singapore. It features an 8192-token context length and utilizes the default Llama-3-8B tokenizer. The model is a key component of the Sahabat-AI ecosystem, co-initiated by GoTo Group and Indosat Ooredoo Hutchison, focusing on Indonesian language and its dialects.

Key Capabilities & Training

• Multilingual Proficiency: Fine-tuned with extensive instruction-completion pairs in Indonesian (448k), Javanese (96k), Sundanese (98k), and English (129k), making it highly capable in these languages.
• Instruction Following: Evaluated using a localized IFEval dataset for Bahasa Indonesia, demonstrating strong adherence to prompt constraints.
• Regional Benchmark Performance: Shows competitive performance on the SEA HELM (BHASA) evaluation benchmark across various tasks (QA, Sentiment, Toxicity, Translation, Summarization, Causal Reasoning, NLI) for Indonesian, Javanese, and Sundanese. It also performs well on IndoMMLU, covering humanities, language, and STEM topics.

Limitations

• Hallucination: Like many LLMs, the model may generate irrelevant or fictional content.
• Reasoning Inconsistencies: Users should validate responses due to potential inconsistencies in reasoning.
• Safety: The model has not been aligned for safety; developers are responsible for implementing their own safety fine-tuning and measures.

Ideal Use Cases

• Applications requiring robust understanding and generation in Indonesian, Javanese, and Sundanese.
• Instruction-following tasks in a multilingual Southeast Asian context.
• Research and development within the Indonesian LLM ecosystem.