Overview

scb10x/llama3.2-typhoon2-1b is a 1 billion parameter decoder-only large language model, developed by scb10x, specifically pretrained for the Thai language. It is built upon the Llama3.2 architecture and supports a context length of 32768 tokens. While primarily focused on Thai, it also supports English.

Key Capabilities

• Thai Language Proficiency: Demonstrates strong performance across various Thai benchmarks, including ThaiExam, ONET, TGAT, and TPAT.
• Llama3.2 Base: Leverages the foundational architecture of Llama3.2-1B.
• Multilingual Support: Capable of processing both Thai and English languages.

Performance Highlights

The model shows competitive results on Thai-specific academic and general knowledge benchmarks:

• ThaiExam: Achieves 26.83%
• TGAT: Scores 49.23%
• TPAT: Reaches 31.03%

Intended Uses & Limitations

This is a pretrained base model, meaning it may require further instruction fine-tuning or one/few-shot learning to follow human instructions effectively. Users should be aware that the model lacks built-in moderation mechanisms and may generate inappropriate or harmful content. For more technical details, refer to the associated arXiv paper.