MyeongHo0621/Qwen2.5-3B-Korean

Overview

MyeongHo0621/Qwen2.5-3B-Korean is a 3.1 billion parameter model, a Korean-optimized version of the Qwen2.5-3B-Instruct base model. Developed by MyeongHo Shin, it has been fine-tuned using 200,000 high-quality Korean conversational data samples, making it highly proficient in Korean language understanding and generation. The model comes with its LoRA adapter already merged, providing a ready-to-use solution for various deployment scenarios.

Key Features

• Korean Optimization: Specifically trained on a substantial dataset of Korean conversational data for enhanced performance in the language.
• Ready-to-Use: Provided as a fully merged model (Safetensors) and in various GGUF quantization formats (Q4_K_M, Q5_K_M, Q8_0, F16) for flexible deployment across different hardware and frameworks like Transformers, vLLM, SGLang, Ollama, and Llama.cpp.
• Commercial Use: Licensed under Apache 2.0, allowing for commercial applications, modifications, and distribution.
• Training Details: Fine-tuned using QLoRA (4-bit NF4) with a LoRA rank of 64 and alpha of 128, over 3 epochs on a dataset of 200,000 Korean dialogue pairs.

Good for

• Korean Conversational AI: Excels in general dialogue, instruction following, and knowledge-based Q&A in Korean.
• Local & Edge Deployment: GGUF formats are ideal for running on local desktops (Ollama) and CPU/edge devices (Llama.cpp).
• Production Serving: Supports vLLM and SGLang for high-throughput inference in production environments.
• Research & Development: A separate PEFT adapter repository is available for further fine-tuning research.