moche-ai/jojangju-KR-31B
jojangju-KR-31B by moche-ai is a 31 billion parameter multimodal language model built on the Gemma-4-31B architecture, specializing in Korean text and image reasoning. It features a 256K token context length and demonstrates significant gains in Korean cultural commonsense (CLIcK) and reasoning tasks (MuSR, Com2) over its base model. This model is optimized for Korean knowledge, commonsense, and reasoning question-answering, as well as multiple-choice evaluations and research.
Loading preview...
Overview
moche-ai/jojangju-KR-31B is a 31 billion parameter multimodal reasoning model, developed by moche-ai, specifically tailored for the Korean language. It is built upon the Gemma-4-31B architecture, inheriting its multimodal capabilities for both text and image processing, and features an extensive 256K token context length. The model utilizes a "merge line" approach, applying task arithmetic on the base Gemma-4-31B model to enhance its performance for Korean-specific tasks.
Key Capabilities & Features
- Korean Multimodal Reasoning: Excels in understanding and generating responses based on both Korean text and images.
- Enhanced Korean Performance: Shows substantial improvements over the base Gemma-4-31B model in Korean cultural/commonsense (CLIcK) and reasoning (MuSR, Com2) benchmarks, with CLIcK scores reaching 0.980 and MuSR at 0.635.
- Proprietary Adaptation: While adaptation details are proprietary, moche-ai confirms that no benchmark evaluation data was used during training or merging to ensure unbiased results.
- High Context Length: Supports up to 262,144 tokens, enabling processing of longer and more complex inputs.
Intended Use Cases
- Korean Knowledge & Commonsense QA: Ideal for question-answering systems requiring deep understanding of Korean culture and general knowledge.
- Reasoning Tasks: Suitable for multi-step and causal reasoning in Korean.
- MCQ Evaluation & Research: Designed for objective evaluation and research purposes, particularly in generative multiple-choice formats.
Limitations
Users should be aware that the model may produce factual errors, hallucinations, or biases, inheriting safety alignment from its base model. Benchmark scores are indicative of performance but do not guarantee real-world quality.