Llama3 8B CPT Sahabat-AI v1 Instruct
| Property | Value |
|---|---|
| Parameter Count | 8.03B |
| Languages | English, Indonesian, Javanese, Sundanese |
| License | Llama3 Community License |
| Context Length | 8192 tokens |
| Base Model | Llama3 |
What is llama3-8b-cpt-sahabatai-v1-instruct-GGUF?
Sahabat-AI v1 Instruct is a multilingual Large Language Model specifically optimized for Indonesian languages. Co-initiated by GoTo Group and Indosat Ooredoo Hutchison, it has been fine-tuned on 448,000 Indonesian instruction pairs, along with 96,000 Javanese and 98,000 Sundanese pairs, plus 129,000 English instruction pairs.
Implementation Details
The model uses the Llama3 architecture with 8B parameters and employs the default Llama-3-8B tokenizer. It has undergone extensive evaluation on multiple benchmarks including SEA HELM, IndoMMLU, and standard English tasks.
- Full parameter fine-tuning completed in 4 hours
- Alignment training conducted for 2 hours
- Trained on 8x H100-80GB GPUs
- Implements on-policy alignment and model merges
Core Capabilities
- Strong performance in Indonesian language tasks (57.221% on BHASA benchmark)
- Effective handling of Javanese (56.460%) and Sundanese (47.495%) content
- Competitive English language capabilities (24.43% average on standard benchmarks)
- 8192 token context length for handling longer conversations
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its specialized optimization for Indonesian languages and dialects, backed by major tech companies and extensive instruction tuning across multiple languages.
Q: What are the recommended use cases?
The model is well-suited for Indonesian language processing tasks, multilingual applications, and general instruction-following scenarios in Indonesian, Javanese, Sundanese, and English contexts.
