phamhai/Llama-3.2-3B-Instruct-Frog

Overview

phamhai/Llama-3.2-3B-Instruct-Frog is a 3.2 billion parameter instruction-tuned model built upon Meta's Llama-3.2-3B-Instruct. Developed by phamhai, this model specifically targets and optimizes performance for Vietnamese Retrieval-Augmented Generation (RAG) tasks. While the base Llama-3.2 models showed limited support for Vietnamese, this fine-tuned version aims to provide robust capabilities for real-world business scenarios where external knowledge supplementation is crucial to avoid hallucinations.

Key Capabilities

• Vietnamese RAG Optimization: Primarily focused on enhancing RAG performance for Vietnamese language applications.
• Efficient Inference: Designed for fast inference, enabling deployment on on-premise and edge devices (laptops, smartphones, NVIDIA Jetson Xavier, Raspberry Pi).
• Extended Context Length: Features a notable 131K context length, allowing for processing longer inputs.
• Function Calling: Demonstrates strong performance in Vietnamese function calling, with the 3B-Instruct-Frog model achieving 95.79% accuracy for function name and 51.05% for exact match on the Vietnamese Function Calling Benchmark.
• Instruction Following: Capable of performing various instruction-based tasks such as Q&A, summarization, and question rewriting, with support for custom personas in system prompts.

Good For

• Vietnamese Language Applications: Ideal for developers building applications that require strong Vietnamese language understanding and generation.
• RAG Systems: Particularly well-suited for integrating with RAG pipelines to provide accurate and contextually relevant responses in Vietnamese.
• Edge Device Deployment: Its optimized performance makes it a viable option for deployment on resource-constrained hardware.
• Function Calling Use Cases: Can be leveraged for applications requiring structured output and interaction with external tools via function calls in Vietnamese.