| Property | Value |
|---|---|
| Parameter Count | 14B |
| Release Date | March 6, 2025 |
| License | CC BY-NC 4.0 |
| Developer | Dnotitia Inc. |
| Languages | Korean, English |
What is DNA-R1?
DNA-R1 is a specialized reasoning model optimized for Korean language, built upon Microsoft's Phi-4 architecture. This innovative model represents a significant advancement in Korean language AI, combining sophisticated reasoning capabilities with deep Korean language understanding. Through a comprehensive three-stage training process, DNA-R1 has been specifically engineered to excel in mathematical reasoning, coding tasks, and general problem-solving while maintaining strong performance in both Korean and English contexts.
Implementation Details
The model employs a sophisticated three-stage training pipeline: initial supervised fine-tuning with 760k Korean examples, strategic integration of reasoning patterns from DeepSeek R1, and advanced GRPO reinforcement learning optimization. This approach has resulted in exceptional performance across various benchmarks, notably achieving 92.49% accuracy on GSM8K and 83.05% on KoBEST.
- Multi-stage training methodology incorporating both supervised and reinforcement learning
- Specialized Korean reasoning dataset integration (300k examples)
- Advanced reward system focusing on format, accuracy, and language consistency
- Comprehensive evaluation across multiple benchmarks showing competitive performance against larger models
Core Capabilities
- Advanced chain-of-thought (CoT) reasoning in Korean
- Self-verification and reflection mechanisms
- Complex problem-solving across mathematics and coding domains
- Cultural and linguistic context maintenance
- Distinct thinking and answer generation using specialized tags
Frequently Asked Questions
Q: What makes this model unique?
DNA-R1 stands out for its specialized optimization for Korean language reasoning, achieved through a unique combination of large-scale Korean datasets and advanced reinforcement learning techniques. Despite being only 14B parameters in size, it demonstrates superior performance compared to many larger models across various benchmarks.
Q: What are the recommended use cases?
The model is particularly well-suited for applications requiring sophisticated reasoning in Korean language contexts, including mathematical problem-solving, coding tasks, and general reasoning challenges. It's especially effective for tasks requiring detailed chain-of-thought reasoning and self-verification capabilities.