SK Telecom announced on the 11th that it has released its independently developed large language model (LLM), 'adot X 3.1 Lite,' to the open-source community 'Hugging Face.'
'adot X 3.1 Lite' was developed by SK Telecom using a 'From Scratch' approach. Data training and other processes were also conducted in-house. It is a lightweight model based on 7 billion (7B) parameters. Leveraging its accumulated LLM technology capabilities, SK Telecom plans to support the government's upcoming 'independent AI foundation model development project.'
'adot X 3.1 Lite' is an upgraded version of the 'adot X 3.0 Lite' model that was applied to the existing adot call summarization. A company spokesperson noted, "It inherits the lightweight performance and high efficiency of its predecessor, and is expected to contribute to corporations' smooth service provision by implementing optimal quality across various specifications of mobile devices regarding information processing capacity and power consumption."
'adot X 3.1 Lite' possesses Korean language processing capabilities on par with its sibling LLM 'adot X 4.0 Lite,' which is based on the same scale of parameters. In the Korean language ability assessment benchmark 'Korean Multitask Language Understanding Evaluation' (KMMLU), 'adot X 3.1 Lite' (61.70) recorded 96% of the score of 'adot X 4.0 Lite' (64.15). In the benchmark for Korean language and culture CLIcK, 'adot X 3.1 Lite' (71.22) outperformed 'adot X 4.0 Lite' (69.97) with a performance level of 102%.
SK Telecom plans to steadily release its LLMs as open-source in the future. The successor to 'adot X 3.1,' a From Scratch model with 34 billion (34B) parameters, will also be announced later this month.
SK Telecom aims to continuously enhance its technological capabilities related to LLMs, including the From Scratch development approach. Since 2018, SK Telecom has been developing AI language models and began utilizing KoBERT in customer center chatbots in 2019. In 2020, it became the first in South Korea to develop and release GPT-2 in Korean as open-source.
The A.(adot) service launched in 2022 incorporated the internally developed adot model. Last year, the adot X 3.0 model was also applied to the adot call summarization feature.
SK Telecom aims to acquire self-sufficient technology as a sovereign AI model with the adot X 3 series while utilizing the adot X 4 series model based on large-scale training (CPT - Continual Pre-Training) in areas that require optimized performance and efficiency. adot X 4 was developed by adding Korean data to the LLM 'Qwen 2.5,' which was released as open-source by Alibaba.
The company stated, "Through the Two Track strategy, we aim to provide domestic corporations with various options to more easily utilize AI technology in their respective environments," and noted that it plans to expand its GPU (graphics processing unit) resources and enhance its development capabilities to significantly innovate the performance of existing models.
Kim Tae-yoon, SK Telecom's foundation model manager, said, "We will strive to enhance the self-sufficiency of the Korean AI ecosystem based on the development capabilities we've steadily built and contribute to improving national AI competitiveness."