EEVE-Korean-Instruct-10.8B-v1.0
| Property | Value |
|---|---|
| Parameter Count | 10.8B |
| License | Apache 2.0 |
| Base Model | SOLAR-10.7B-v1.0 |
| Paper | Research Paper |
What is EEVE-Korean-Instruct-10.8B-v1.0?
EEVE-Korean-Instruct-10.8B-v1.0 is an advanced Korean language model that builds upon the SOLAR-10.7B architecture with specialized Korean vocabulary extensions. The model was fine-tuned using Direct Preference Optimization (DPO) through Axolotl, incorporating translated versions of high-quality datasets including SlimOrca-Dedup and UltraFeedback.
Implementation Details
The model utilizes a BF16 tensor type and implements a specific prompt template for chat-based interactions. It achieves impressive scores on various benchmarks, including 83.04 on HellaSwag and 81.93 on Winogrande, demonstrating strong performance in both Korean and cross-lingual tasks.
- Korean-optimized vocabulary extension
- DPO fine-tuning methodology
- Comprehensive benchmark performance
- Efficient implementation using Transformers architecture
Core Capabilities
- Advanced Korean language understanding and generation
- Multi-turn conversational abilities
- Strong performance on reasoning and knowledge tasks
- Efficient processing of both Korean and English inputs
Frequently Asked Questions
Q: What makes this model unique?
The model's unique strength lies in its specialized Korean vocabulary extension and careful fine-tuning using DPO, making it particularly effective for Korean language tasks while maintaining strong general language capabilities.
Q: What are the recommended use cases?
The model is best suited for Korean language applications including conversational AI, content generation, and complex reasoning tasks. It's particularly effective in scenarios requiring detailed and polite responses in Korean.
