Karasu-DPO-7B
| Property | Value |
|---|---|
| Parameter Count | 7 Billion |
| Context Length | 1024 tokens |
| Language | Japanese |
| License | Apache 2.0 |
| Base Model | Qwen2.5-7B-Instruct |
| Model URL | lightblue/Karasu-DPO-7B |
What is Karasu-DPO-7B?
Karasu-DPO-7B is a specialized Japanese language model that builds upon the Qwen2.5-7B-Instruct architecture. It's been specifically optimized for Japanese conversation through Direct Preference Optimization (DPO) training using synthetic Japanese dialogue data. The model demonstrates significant improvements over its base version, achieving a 66.2% score on the arena-hard-auto-multilingual chat benchmark compared to the base model's 50.0%.
Implementation Details
The model was developed using a sophisticated training procedure that involved sampling from multiple high-quality datasets, including lmsys-chat-1m, ShareGPT52K, and OpenAssistant/oasst2. The training process utilized QLoRA DPO with carefully tuned hyperparameters: learning rate of 5e-6, batch size of 4, and cosine learning rate scheduling.
- Synthetic data generation using GPT-4
- Multi-stage translation and correction process
- QLoRA DPO fine-tuning methodology
- Comprehensive validation process with documented loss metrics
Core Capabilities
- Advanced Japanese language understanding and generation
- Optimized for conversational AI applications
- Supports both casual and technical discussions
- Efficient deployment through vLLM integration
Frequently Asked Questions
Q: What makes this model unique?
The model's unique strength lies in its specialized Japanese language capabilities, achieved through a careful DPO training process using high-quality synthetic data. The significant performance improvement over the base model makes it particularly valuable for Japanese-language applications.
Q: What are the recommended use cases?
Karasu-DPO-7B is primarily recommended for general conversation AI applications in Japanese. It's particularly well-suited for deployments requiring natural Japanese language interaction, whether in customer service, content generation, or interactive applications.
