Biawak-8B-Base is an 8-billion-parameter causal language model developed by AITF Indonesia, built upon the Qwen-3-8B architecture. This model is specifically adapted through continued pre-training on a curated Indonesian dataset, focusing on Digital Space Protection (PRD) and Digital Talent Pool (DTP) domains. It excels in understanding Indonesian digital policies, cybersecurity, and workforce development, making it suitable for specialized text completion tasks within these areas.
Loading preview...
Biawak-8B-Base: A Domain-Specialized Indonesian LLM
Biawak-8B-Base is an 8-billion-parameter Large Language Model (LLM) developed by AITF Indonesia. It is built on the Qwen-3-8B base model through Continued Pre-training (CPT), specifically adapted for Indonesia's strategic focus areas: Digital Space Protection (PRD) and Digital Talent Pool (DTP).
Key Capabilities & Training
- Domain Specialization: Trained on a 214.2 million token Indonesian dataset, with significant portions dedicated to PRD (42.9%) and DTP (~43.9%) topics, alongside general Indonesian Wikipedia data.
- Language Focus: Primarily Indonesian, with secondary English language support.
- Base Model: Functions as a base causal language model, designed for text completions and adaptable into chat/instruct variants through further fine-tuning.
- Training Hardware: Continued pre-training was conducted on NVIDIA A100 80GB GPUs for approximately 36 hours.
Intended Use Cases
This model is designed to provide a sovereign, domain-specialized Indonesian foundation model with strong understanding in:
- Digital Space Protection (PRD):
- Policy sentiment analysis
- Misinformation pattern detection
- Understanding legal terminology (e.g., UU ITE, UU PDP)
- Digital Talent Pool (DTP):
- Skill gap analysis
- Curriculum drafting assistance
- Job description and talent understanding
Limitations and Recommendations
As a base model, Biawak-8B-Base requires Supervised Fine-Tuning (SFT) for optimal performance in specific PRD/DTP applications. Users should be aware of potential biases inherited from web data and the possibility of factual hallucinations. It is recommended to add high-quality instruction datasets and perform evaluation benchmarks before production deployment.