The on-device artificial intelligence (AI) model of LG AI Research Institute (President Bae Kyung-hoon) ranked first in the sting section of Hugging Face’s ‘Large Language Model (LLM) Leaderboard’.
As of the nineteenth, ‘ExaOne-3.5-2.4B-Instruction’, released as open source by LG AI Research Institute, recorded the best rating within the leaderboard edge category, beating Google’s ‘Gemma 2’ and Alibaba’s ‘Q1 2.5’. did it
This model is the smallest among the many ‘ExaOne 3.5’ model group released on the tenth. At the moment, along with the two.4B, the 7.8B model and the frontier-class 32B were released.
Even on the time of release, it attracted attention with its excellent benchmark performance. It was revealed that it showed superior performance to global big tech’s flagship open source models comparable to Gemma 2 and Q1 2.5, in addition to Meta’s ‘Rama 3.1’ and Microsoft’s ‘Pi-3’.
And the Hugging Face LLM leaderboard is a stage for evaluating models which have improved performance by fine-tuning these foundation models. Here too, ExaOne still achieved superior results.
Specifically, it recorded 79.5 points and 34 points in IFEVAL (instruction ability) and MATH (mathematics) fields, respectively, surpassing competing models by an amazing margin.
It is usually noteworthy that it achieved the very best leads to the lightweight model category. Recently, fierce competition is happening in the sector of edge small language models (sLM) in addition to in frontier models.
Evidence is that in recent days, starting with Antropic, MS, Cohere, and UAE’s TII have launched sLM one after one other. On today, IBM launched the 2B model, and at one point it rose to third place on the leaderboard.
As well as, the ExaOne-7.8B model ranked 261st overall. This can be a results of raising the rating by as much as 800 places in comparison with the previous version. The biggest, ExaOne-32B, isn’t listed on the leaderboard.
An LG official said, “We’re joyful to have achieved a meaningful achievement of being first within the leaderboard edge category. We is not going to rest on our laurels, but will proceed to match performance and reflect it in our next model research.”
Meanwhile, that is the second time that a domestic model has ranked first because the launch of Hugging Face Leaderboard Season 2 in June. On the eighth of last month, data specialist N3N (CEO Nam Young-sam) announced that it ranked first on the planet within the 11B and below category.
Reporter Jang Se-min semim99@aitimes.com
