Saxo/Linkbricks-Horizon-AI-Korean-Mistral-Nemo-sft-dpo-12B Overview

Developed by Linkbricks Horizon-AI's data scientist Yunsung Ji (Saxo), this 12 billion parameter model is a Korean language model fine-tuned from the Mistral-Nemo-Instruct-2407 base. The fine-tuning process involved Supervised Fine-Tuning (SFT) followed by Direct Preference Optimization (DPO) using four H100-80G GPUs on KT-CLOUD.

Key Capabilities

• Multilingual Cross-Augmentation: Trained with Korean, Chinese, English, and Japanese cross-training data, alongside logical data, to improve multilingual understanding and logical reasoning.
• Complex Korean Logic: Specifically designed to address and solve intricate logical problems in Korean.
• Enhanced Analysis: Strengthened for high-level analysis of customer reviews and social media postings.
• Coding Proficiency: Demonstrates enhanced capabilities in coding tasks.
• Extended Context Window: Features a substantial context window size of 128K tokens, allowing for processing longer inputs.
• Tokenizer Consistency: Utilizes the base model's tokenizer without word expansion.

Performance & Training Details

• Achieved Rank-4 on the Open Ko LLM Leaderboard Season 2 as of November 1, 2024.
• Training employed Deepspeed Stage 3 and rslora techniques.

Good For

• Applications requiring advanced Korean language understanding and logical problem-solving.
• Analyzing customer feedback and social media content in Korean.
• Coding assistance and generation in a multilingual context.
• Use cases benefiting from a large context window for processing extensive text.