Overview

scb10x/llama3.1-typhoon2-70b-instruct is a 70 billion parameter instruction-tuned large language model, built upon the Llama3.1 architecture. Developed by scb10x, this model is primarily focused on Thai language capabilities while also demonstrating strong performance in English. It features a notable 90k context length, allowing for deep contextual understanding and generation.

Key Capabilities

• Bilingual Proficiency: Excels in both Thai and English, particularly in instruction-following and function calling.
• Domain-Specific Performance: Shows strong results in mathematical reasoning and coding tasks in both languages.
• Long Context Handling: Designed to process and generate content with a context length of up to 90,000 tokens.
• Function Calling: Demonstrates robust function calling capabilities, with higher performance in Thai (70.8%) compared to English (65.7%).

Performance Highlights

Compared to Llama3.3 70B Instruct and Openthaigpt1.5 72B, Typhoon2 Llama3.1 70B Instruct achieves leading scores in:

• IFEval - TH: 81.45%
• MT-Bench TH: 7.3626
• Thai Code-Switching: 98.8% (t=0.7) and 94.8% (t=1.0)
• GSM8K - TH: 88.79%
• MATH - TH: 59.60%

Good For

• Applications requiring high-quality Thai language generation and understanding.
• Instruction-following tasks and function calling in a bilingual (Thai/English) environment.
• Use cases demanding long context processing, such as document analysis or extended conversations.
• Developing AI solutions for math and coding challenges, especially with a focus on Thai language support.