AXCXEPT/phi-4-deepseek-R1K-RL-EZO
AXCXEPT/phi-4-deepseek-R1K-RL-EZO is a 14.7 billion parameter language model developed by AXCXEPT, combining the Phi-4 architecture with reinforcement learning inspired by Deepseek R1 research. It features enhanced multilingual performance in both Japanese and English, outperforming the base Phi-4 model and gpt-4o-mini in various benchmarks. This model is optimized for high-security enterprise applications requiring local LLM deployment and strict data privacy compliance.
Loading preview...
AXCXEPT/phi-4-deepseek-R1K-RL-EZO Overview
This model, developed by AXCXEPT, integrates the Phi-4 architecture with a novel reinforcement learning (RL) approach, drawing insights from Deepseek R1 research. The primary goal was to enhance both Japanese and English language capabilities while maintaining strong overall performance. The model was fine-tuned using a 14K dataset in just two days, demonstrating efficient training.
Key Capabilities & Improvements
- Enhanced Multilingual Performance: Strengthens English capabilities without compromising Japanese proficiency, a notable improvement over previous iterations.
- Optimized Training Efficiency: Achieved significant gains through a fine-tuning process inspired by Deepseek R1, completed rapidly.
- Benchmark-Proven Quality: Outperforms the base Phi-4 model on OpenAI’s Simple-eval and translation benchmarks (Japanese MT Bench, MT Bench). It also surpasses gpt-4o-mini in multiple evaluation categories, positioning it as a high-performance 14B model.
Why Local LLMs Matter
This model is specifically designed for enterprises requiring high security and strict data privacy compliance, where cloud-based models are not suitable. It caters to organizations in public institutions, manufacturing, and design industries that need state-of-the-art performance within a secure, closed environment.
Future Prospects
The successful short-term training experiment highlights the potential for developing domain-specific LLMs tailored for high-security industries. AXCXEPT plans to continue refining this methodology and creating specialized AI models for enterprise applications, including SaaS offerings, to accelerate LLM adoption in Japan and globally.