Hugging Face The denial07/Qwen2-72B-Instruct-kor-dpo model is an instruction-tuned large language model with 72.7 billion parameters and a 131,072-token context length, based on the Qwen2-72B-Instruct architecture. This version is specifically improved for Korean language performance, demonstrating enhanced capabilities across various Korean benchmarks. It is optimized for general instruction-following tasks in Korean, including reasoning, math, writing, and coding.
Overview
This model, denial07/Qwen2-72B-Instruct-kor-dpo, is an enhanced version of the Qwen2-72B-Instruct model, specifically fine-tuned to improve its performance in the Korean language. It features 72.7 billion parameters and supports a substantial context length of 131,072 tokens, making it suitable for complex and lengthy Korean language tasks.
Key Capabilities and Performance
The model demonstrates strong performance across several Korean-specific benchmarks:
- LogicKor Benchmark: Achieves a total score of 8.79, ranking competitively among large language models. It shows notable strengths in Coding (9.71), Understanding (9.64), and Writing (9.14) within the Korean context, outperforming the base Qwen2-72B-Instruct model in Reasoning, Math, Writing, and overall score.
- KMMLU Benchmark: Maintains an overall accuracy of 0.58, performing on par with the base Qwen2-72B-Instruct and slightly better than Mistral-Large-Instruct-2407 across categories like HUMSS, STEM, Applied Science, and Other.
When to Use This Model
This model is particularly well-suited for applications requiring high-quality Korean language understanding and generation. Its improvements over the base Qwen2 model make it a strong candidate for tasks such as:
- Korean-centric content creation and summarization.
- Complex reasoning and problem-solving in Korean.
- Code generation and understanding for Korean-speaking developers.
- Instruction-following tasks where nuanced Korean language comprehension is critical.