Hunminai-1.0-12b: Korean-Aligned Gemma-3 Model
Hunminai-1.0-12b is a 12 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 on Korean natural language tasks, utilizing a corpus of 100,000 instruction examples through Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO).
Key Capabilities & Features
- Korean Language Alignment: Optimized for understanding and generating Korean text, aligning with user intents.
- Fine-Tuning: Employs SFT and DPO techniques for improved performance and alignment.
- Versatile Applications: Suitable for dialogue generation, question answering, and long-form text generation in Korean.
- Strong Benchmark Performance: Achieves an average score of 7.80 across a suite of Korean LLM evaluation benchmarks, outperforming its base model and several other Korean-focused models in various categories like GPQA and IFEVAL.
Evaluation Highlights
The model was rigorously evaluated on a range of Korean benchmark datasets, including ko-bench, ko-ged (elementary, middle, high school), ko-ifeval, ko-gpqa, and ko-math-500. Notably, Hunminai-1.0-12b scored 9.72 on ged:E (Korean elementary school GED multiple-choice), 3.18 on gpqa (challenging physics questions), and 8.37 on ifeval (instruction-following evaluation).
Good for
- Developing AI assistants and chatbots for Korean-speaking users.
- Applications requiring high-quality Korean text generation.
- Research and development in Korean natural language processing.