typhoon-ai/llama3.1-typhoon2-8b-instruct
Llama3.1-Typhoon2-8B-instruct is an 8 billion parameter instruction-tuned large language model developed by typhoon-ai, based on the Llama3.1 architecture. Optimized for bilingual performance, it excels in Thai and English instruction-following, function calling, and specific domains like math and coding. With a 90k context length, it is designed for applications requiring robust performance in both languages.
Loading preview...
Llama3.1-Typhoon2-8B-instruct: A Bilingual Thai-English LLM
Llama3.1-Typhoon2-8B-instruct is an 8 billion parameter instruction-tuned model built upon the Llama3.1 architecture, developed by typhoon-ai. This model is specifically designed to offer strong performance in both Thai (🇹🇭) and English (🇬🇧) languages, making it a versatile choice for bilingual applications.
Key Capabilities & Performance
- Bilingual Instruction Following: Demonstrates significantly improved performance in Thai instruction following (72.60% on IFEval-TH) compared to the base Llama3.1 8B Instruct (58.04%), while maintaining strong English performance.
- Enhanced Function Calling: Achieves 75.12% in Thai FunctionCall and 79.08% in English FunctionCall, indicating robust tool-use capabilities.
- Domain-Specific Strengths: Shows notable improvements in Thai math (GSM8K-TH 71.72%, MATH-TH 38.48%) and coding (HumanEval-TH 58.5%) benchmarks.
- Long Context Handling: Features an extended context length of 90,000 tokens, enabling processing of extensive inputs and generating comprehensive responses.
- Code-Switching Proficiency: Excels in Thai Code-Switching tasks, achieving 98.8% at t=0.7 and 98% at t=1.0.
When to Use This Model
This model is particularly well-suited for use cases requiring high-quality instruction following and function calling in a bilingual Thai-English context. Its strong performance in specific domains like math and coding, combined with its long context window, makes it ideal for applications such as:
- Bilingual chatbots and virtual assistants.
- Content generation and summarization in Thai and English.
- Code generation and problem-solving for developers working with both languages.
- Applications requiring complex function calling and tool integration.
It is important to note that while the model incorporates guardrails, it is still under development and may produce inaccuracies or biases. Developers should assess these risks for their specific use cases. For more technical details, refer to the arxiv paper.