Hunminai-1.0-27b: Korean-Aligned Gemma-3 Model
Hunminai-1.0-27b is a 27 billion parameter language model developed by davidkim205, built upon Google's Gemma-3 architecture. This model has been specifically fine-tuned to enhance its performance and alignment with Korean natural language tasks. The fine-tuning process involved Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) on a corpus of 100k Korean instruction examples.
Key Capabilities and Features
- Korean Language Optimization: Designed to excel in Korean dialogue generation, question answering, and long-form text generation.
- Advanced Fine-Tuning: Utilizes SFT and DPO to better align with user intents in Korean.
- Gemma-3 Base: Leverages the robust architecture of
google/gemma-3-27b-it. - Extended Context Length: Supports a context length of 128k tokens.
Performance Highlights
Evaluated against a suite of Korean LLM benchmarks, Hunminai-1.0-27b consistently shows strong performance:
- Achieves an average score of 8.51, surpassing
google/gemma-3-27b-it (8.32) and unsloth/gemma-3-27b-it (8.03). - Demonstrates superior results in specific benchmarks like
gpqa (4.55 vs. 3.69 for base Gemma-3) and math500 (8.56 vs. 8.38 for base Gemma-3).
Ideal Use Cases
This model is particularly well-suited for applications requiring high-quality, instruction-following capabilities in Korean, such as:
- Developing Korean-language chatbots and virtual assistants.
- Generating coherent and contextually relevant long-form Korean text.
- Implementing advanced Korean question-answering systems.