Overview

kiki-ailab/Llama3.2-3B-Instruct-KAI is a 3.2 billion parameter instruction-tuned model, fine-tuned by kiki-ailab from the Llama3.2 base. It is part of a collection of models optimized for Vietnamese language tasks.

Key Capabilities

• Vietnamese Language Understanding: Excels in processing and generating Vietnamese text.
• Instruction Following: Designed to respond effectively to instructions for various NLP tasks.
• Task Performance: Optimized for reading comprehension, information extraction, question answering, and summarization in Vietnamese.

Performance Highlights

The model shows substantial gains over its base Llama3.2-3B-Instruct on several Vietnamese benchmarks:

• VMLU: Achieves 58.1, an improvement of +10.5 points.
• ViSquad: Scores 93.5, up +3.2 points.
• ViDrop: Reaches 81.4, a significant +17.9 point increase.
• ViDialog: Attains 67.3, an improvement of +16.5 points.

On the ArenaHard benchmark, kiki-ailab/Llama3.2-3B-Instruct-KAI demonstrates competitive performance among 3B-sized models, outperforming the original Llama-3.2-3B-Instruct with a 35% win rate.

Limitations

• May exhibit hallucinations, particularly with cultural-specific content.
• Primarily focused on Vietnamese language; performance in specialized technical domains may not be optimal.