AXCXEPT/EZO2.5-gemma-3-12b-it-Preview

AXCXEPT/EZO2.5-gemma-3-12b-it-Preview is a 12 billion parameter instruction-tuned model developed by AXCXEPT, based on Google's Gemma-3 architecture. It leverages a proprietary "EZO" training method, integrating GRPO and PPO concepts, to significantly enhance Japanese language performance on benchmarks like Japanese MT Bench and Elyza Tasks100. This model is optimized for improving base model capabilities with limited data and computational resources, offering a cost-effective alternative to complex reinforcement learning.