Overview of DimensionSTP/gemma-3-12b-it-Ko-Reasoning
This model is a 12 billion parameter, instruction-tuned variant of Google's Gemma-3-12b-it, developed by DimensionSTP. It is specifically engineered to excel in logical and multi-hop reasoning tasks in Korean, addressing a critical need for specialized reasoning capabilities in the language. The fine-tuning process involved a comprehensive Korean-English instruction dataset, incorporating diverse multi-hop questions, symbolic logic tasks, and human-crafted reasoning steps.
Key Capabilities & Differentiators
- Korean Reasoning Specialization: Optimized for complex logical and multi-hop reasoning within the Korean language context.
- Enhanced from Gemma-3-12b-it: Builds upon the robust foundation of Google's Gemma 3 architecture.
- Extensive Context Window: Features a substantial context length of 128,000 tokens, beneficial for intricate reasoning problems.
- Benchmark Performance: Demonstrates strong performance on various reasoning benchmarks, including GPQA diamond (61.3%), GSM8K (59.6%), HAERAE (73.9%), KSM (66.7%), LogicKor (8.56), and Math500 (77.8%), all measured using 0-shot Chain-of-Thought (CoT).
- Open-Source Initiative: Part of the Ko-Reasoning Series, aiming to provide open-access models that can rival proprietary solutions in complex reasoning.
Ideal Use Cases
This model is particularly well-suited for applications requiring advanced reasoning and problem-solving in Korean, such as:
- Complex Question Answering: Handling multi-step questions that require logical deduction.
- Symbolic Logic Tasks: Solving problems based on symbolic reasoning.
- Educational Tools: Assisting with Korean-language math and logic problems.
- Research and Development: Exploring the boundaries of open-source reasoning models in Korean.