FINAL-Bench/JGOS-31B-Citizen-R
FINAL-Bench/JGOS-31B-Citizen-R is a 31 billion parameter Gemma4-based model developed by FINAL-Bench, specifically optimized for Korean multi-step reasoning. It utilizes a unique "Darwinian cross-breeding" technique, transplanting Feed-Forward Network (FFN) layers from various models into the original Citizen backbone to enhance Korean reasoning capabilities across diverse domains. This model excels in complex inference tasks, particularly when activated in its designated "thinking mode," and supports a context length of 32768 tokens.
Loading preview...
JGOS-31B-Citizen-R: Korean Multi-Step Reasoning Model
JGOS-31B-Citizen-R is a 31 billion parameter model based on the Gemma4 architecture, developed by FINAL-Bench. It is specifically designed and optimized for advanced multi-step reasoning in Korean. This model significantly enhances the capabilities of its base model, JGOS-31B-Citizen, across various Korean-specific benchmarks.
Key Capabilities
- Enhanced Korean Multi-Step Reasoning: Achieves substantial improvements in multi-step narrative reasoning (MuSR(Ko)).
- Broad Knowledge & Common Sense: Shows concurrent gains across knowledge, common sense, cultural, and linguistic subjects, as evidenced by improved performance on KMMLU, Com2, and CLIcK benchmarks.
- Optimized "Thinking Mode": Performance is maximized when explicitly invoked in its "thinking mode," which activates strengthened Korean inference pathways through FFN transplantation.
Good For
- Complex Korean Inference Tasks: Ideal for applications requiring deep, multi-layered reasoning in the Korean language.
- Knowledge-Intensive Korean NLP: Suitable for tasks involving extensive Korean knowledge, common sense, and cultural understanding.
- Applications Requiring High Context: Supports a substantial context length of 32768 tokens, beneficial for detailed analysis and long-form content processing.