Deepnoid (CEO Choi Woo-sik), Korea's first-generation medical artificial intelligence (AI) specialist, won the 'Open Ko-LLM Leaderboard' hosted by the Korea Intelligence and Information Society Agency (NIA) and Upstage. It was announced on the fifteenth that the model ‘deep solar (deep-solar-v2.0.1)’ ranked first.
This model was created based on Upstage's large language model (LLM) 'Solar'.
Specifically, it recorded an overwhelmingly high rating of 70 points in 'Ko-ARC (reasoning ability)', which measures how appropriate the AI's answers to questions are.
The improved 'Reinforcement Learning through Human Feedback (RLHF)' was chosen as a feature. To forestall overfitting that happens during data learning, regularization might be added to the loss to extend the steadiness of learning.
Hyun Ji-hoon, head of Deepnoid AI Research Center, said, “Deepnoid has already researched and developed image reading solutions using AI in various fields and is producing results,” adding, “We’ll showcase more advanced solutions and technologies by connecting Deep-Solar with various businesses.” “He said.
Meanwhile, this model recorded a median rating of 61.45 points, raising the common by about 1 point per week after the 60-point barrier on the Ko-LLM leaderboard was broken. As such, performance improvement is predicted to speed up within the Korean model leaderboard that began in September of last 12 months.
Reporter Jang Se-min semim99@aitimes.com
