dnotitia/DNA-R1
DNA-R1 is a 14.7 billion parameter reasoning model developed by Dnotitia Inc., based on Microsoft's Phi-4 architecture, and optimized for Korean language tasks. It significantly enhances Korean reasoning capabilities across mathematics, coding, and general reasoning through large-scale reinforcement learning, using a methodology similar to DeepSeek-R1. The model supports both Korean and English, excelling in generating nuanced chains-of-thought and solving complex multi-step problems while maintaining cultural and linguistic context.
Loading preview...
Overview
dnotitia/DNA-R1 is a 14.7 billion parameter reasoning model developed by Dnotitia Inc., built upon Microsoft's Phi-4 architecture. It is specifically optimized for Korean language reasoning, leveraging a multi-stage training pipeline that includes large-scale reinforcement learning (RL) inspired by DeepSeek-R1's methodology. The model demonstrates advanced reasoning abilities in Korean across mathematics, coding, and general reasoning tasks, supporting both Korean and English languages.
Key Capabilities
- Specialized Korean Reasoning: Enhanced understanding and reasoning depth for Korean text, including self-verification and reflection.
- Multi-Stage Training: Utilizes a three-stage pipeline: initial supervised fine-tuning (SFT) with a large Korean non-reasoning dataset, integration of Korean reasoning patterns from DeepSeek R1, and advanced reinforcement learning with GRPO for format, accuracy, and language consistency.
- Chain-of-Thought (CoT) Generation: Capable of generating sophisticated and nuanced Korean chains-of-thought for complex problem-solving.
- Contextual Understanding: Maintains cultural and linguistic context in reasoning, distinguishing between deep thinking and concise answers using
<think>and<answer>tags.
Performance Highlights
DNA-R1 shows strong performance across various benchmarks, often outperforming larger models. Notable scores include 92.49 on GSM8K, 89.4 on Math500, and competitive results on KMMLU and KoBEST, demonstrating its effectiveness in Korean-specific and general reasoning tasks.