prototie/prototie-ai-final

Overview

prototie/prototie-ai-final is a 14 billion parameter Korean AI model built upon the Qwen3-14B architecture. Developed by Prototie, this model is specifically engineered for enhanced Korean reasoning capabilities through a unique three-stage training pipeline.

Key Capabilities

• Specialized Korean Reasoning: The model is fine-tuned to excel in complex reasoning tasks within the Korean language context.
• Advanced Training Pipeline: It utilizes a sophisticated three-stage training process:
  • STEP1: SFT (Supervised Fine-Tuning): Initial fine-tuning on datasets like KMMLU, Korean mathematics, and CLIcK data.
  • STEP2: GRPO (Generalized Reinforcement Learning from Human Feedback): Reinforcement learning with a triple reward function focusing on accuracy, Korean language quality, and format adherence.
  • STEP3: DARE-TIES: A model merging technique combining the GRPO model (with a weight of 0.7) and the SFT model (with a weight of 0.3) to achieve optimal performance.

Good For

• Applications requiring strong Korean language understanding and generation.
• Tasks involving Korean reasoning and problem-solving.
• Use cases where linguistic accuracy and contextual understanding in Korean are critical.