SEOKDONG/llama3.1_korean_v1.1_sft_by_aidx

SEOKDONG/llama3.1_korean_v1.1_sft_by_aidx: Korean-Specialized Llama 3.1

This 8 billion parameter model, developed by SEOKDONG, is a supervised fine-tuned (SFT) version of Llama 3.1 8B Instruct, specifically optimized for the Korean language, culture, and enterprise domains. It has demonstrated strong performance, ranking #1 among Korean-developed models and #6 overall out of 1,228 models on the Open Ko-LLM Leaderboard Season 2 (NIA, January 2025), competing effectively with larger 14B–27B models.

Key Capabilities & Features

• Korean Language & Culture Specialization: Fine-tuned on a self-built dataset of ~3.6 GB / 2.33M examples, covering 53 domains with a focus on Korean social values and professional contexts.
• Enterprise Domain Expertise: Strong performance in legal, financial, tax, and accounting tasks, including Q&A and document summarization.
• Competitive Performance: Achieved an average score of 53.94 on the Open Ko-LLM Leaderboard S2, outperforming many larger models in its category.
• Chain-of-Thought Formatting: Training data includes 1.33M multiple-choice items formatted with Chain-of-Thought for enhanced reasoning.

Good For

• Enterprise Applications: Ideal for legal, financial, tax, and accounting Q&A and document summarization in Korean.
• Education: Generating Q&A and explanations across Korean history, mathematics, and science.
• Research & Culture: Korean-context NLP, sentiment analysis, and document generation.
• Customer Service: Developing conversational AI for Korean-speaking users.

Limitations

While highly specialized for Korean, its accuracy may decrease for other languages, very recent international information, or highly niche technical fields. It may also exhibit limitations in complex logical reasoning and potential biases from its training data.