OLAIR ko-r1-7b-v2.0.3
| Property | Value |
|---|---|
| Model Size | 7B parameters |
| Type | Korean Language Model |
| Author | OLAIR |
| Model URL | HuggingFace |
| Contact | spthsrbwls123@yonsei.ac.kr |
What is ko-r1-7b-v2.0.3?
ko-r1-7b-v2.0.3 is a specialized Korean language model designed for advanced reasoning tasks. Built by OLAIR, this model represents version 2.0.3 in their series, featuring significant improvements in Korean language understanding and reasoning capabilities through refined training methodologies.
Implementation Details
The model was trained on the Open-R1-Ko-SFT-v2.0 dataset, specifically curated for supervised fine-tuning to enhance Korean language reasoning abilities. Performance evaluations were conducted using the HAE-RAE Reasoning Challenge (HRC), covering multiple domains including mathematics, physics, chemistry, and puzzle-solving.
- Enhanced mathematical reasoning with 61.82% accuracy on math tasks
- Improved performance in physics word puzzles (40% accuracy)
- Overall average performance of 29.94% across all domains
- Noticeable improvements from version 1.0
Core Capabilities
- Specialized Korean language understanding
- Strong performance in mathematical reasoning tasks
- Moderate capability in physics-related problems
- Structured problem-solving across various domains
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its specialized focus on Korean language reasoning tasks, particularly excelling in mathematical problems while maintaining balanced performance across other domains. It represents a significant step forward in Korean language AI capabilities.
Q: What are the recommended use cases?
The model is best suited for Korean language applications requiring mathematical reasoning and problem-solving capabilities. However, users should note its current limitations, particularly in chemistry-related tasks and certain aspects of puzzle-solving.
Q: Are there any known limitations?
Yes, the model currently shows vulnerability to certain Korean-related inputs, which can lead to endless loops of thinking. The development team is actively working on addressing these issues.
