websfactory/Webs-Sejong-31B
Webs-Sejong-31B is a 31-billion parameter Korean-centric language model developed by websfactory, built on the Gemma-4 architecture through weight-space model merging. It integrates the JGOS-31B-Citizen model to enhance Korean cultural and academic competence while retaining English ability. Ranked 3rd on the NIA K-AI Leaderboard, this merge model is optimized for Korean-language assistance, knowledge QA, and reasoning tasks.
Loading preview...
Webs-Sejong-31B: Korean-Centric Merge Model
Webs-Sejong-31B is a 31-billion parameter language model from websfactory, created by weight-space merging the instruction-tuned Gemma-4-31B foundation with the top-ranked open Korean model, JGOS-31B-Citizen. This model is explicitly a merge, not a separately trained entity, and utilizes a sign-consensus task-vector merge (DARE-TIES) to combine their strengths.
Key Capabilities & Features
- Korean-First Optimization: Designed with a strong focus on Korean cultural knowledge and professional/academic reasoning, while maintaining English language proficiency inherited from its base model.
- Verified Performance: Achieved 3rd place on the NIA K-AI Leaderboard with an average score of 0.611 on a private Korean benchmark suite (KMMLU-Pro, CLIcK, HLE, MuSR, Com2), demonstrating independently verified competence.
- Standard Compatibility: Built on the Gemma-4 architecture and tokenizer, ensuring seamless integration with standard
transformersand vLLM serving without requiring custom code. - Merge Method Transparency: The model openly states its creation via DARE-TIES merging, combining
google/gemma-4-31B-itwithJGOS-Model/JGOS-31B-Citizenwithout additional fine-tuning.
Intended Use Cases
- Korean-language Assistance: Ideal for applications requiring strong understanding and generation in Korean.
- Knowledge QA: Suitable for question-answering systems focused on Korean cultural and academic domains.
- Reasoning Tasks: Can be applied to tasks requiring professional and academic reasoning in Korean contexts.