7/19/2023

• Upstage ranked #1 in the HuggingFace Open LLM Leaderboard

• Upstage 30B LLM beat larger models of tech giants, solidifying its leadership in the private AI market

• PRIVATE LLM PRESENTS THE BEST ALTERNATIVE TO COMPANIES' BAN ON CHATGPT DUE TO SECURITY AND HALLUCINATION PHENOMENON.

(Seoul, Jul. 19, 2023 /Upstage) Upstage, South Korea’s leading AI startup, has released the world's most advanced open-source generative AI model.

Upstage announced on July 19 that the company’s 30-billion-parameter Large Language Model (LLM) secured the top position on HuggingFace's Open LLM Leaderboard.

Hugging Face, with its mission to advance and democratize artificial intelligence through open source and open science, offers an evaluation tool that ranks and evaluates open-source LLMs and chatbots. Their LLM leaderboard boasts over 300 models tailored for various languages and purposes, comparable in capabilities to OpenAI's ChatGPT and Google's Bard, where numerous models being submitted each day. Evaluation is based on the average of four key metrics, such as the model’s reasoning challenge, common sense inference, contextual understanding, and factual accuracy.

LLM has garnered significant attention as the pinnacle of generative AI with its capacity to train on a plethora of text data and perform NLP (natural language processing)-based tasks. Major tech companies, including OpenAI's ChatGPT, Google's Bard, and Meta's LLaMa, have already jumped into the fierce competition in this field, while open-source LLMs are also making remarkable progress with models under 100 billion parameters via open-source platforms such as HuggingFace.

UPSTAGE SUBMITTED ITS SELF-BUILT MODEL TO HUGGING FACE'S LEADERBOARD EARLY THIS MONTH AND HAD ITS PERFORMANCE EVALUATED. AS A RESULT, UPSTAGE'S MODEL RANKED SECOND WITH AN AVERAGE SCORE OF 64.7 POINTS, FOLLOWING META'S 'RAMA 2' 70B MODEL, WHICH WAS UNVEILED BY ZUCKERBERG THIS MORNING. THIS IS THE HIGHEST RANKING OF THE 30B (30 BILLION) PARAMETER MODEL, PRODUCING A COMPETITIVE RESULT WITH LESS THAN HALF THE SIZE OF LLAMA 2.

In particular, Upstage's model recorded 56.5 points in the hallucination prevention index, which is one of the biggest problems with generated AI, despite its relatively small model size, a surprising result that far exceeds the 52.8 point rating of Meta's latest 'Rama 2' model. recorded. This model includes models created by big tech companies such as Meta and MS, excluding Lamar2, and models from leading global AI/LLM companies such as Stability AI and Databricks, and is the UAE Technology Innovation Research Institute's '', which has maintained first and second place in recent months. It ranked second with an average of 10% higher performance than the highest-performing AI models, such as the 'Falcon' model and the model of MosaicML, which Databricks recently acquired for $1.3 billion (approximately 1.7 trillion won).

In particular, Upstage surprised everyone by being known to have created the world's highest-performance AI model in about two months after starting to build its own model. Upstage not only built 'KLUE', the first Korean natural language understanding (NLU) evaluation dataset, but also operates Korea's leading generated AI service AskUp, which has surpassed 1.3 million users, and is the best prompt engineering and technology company in the country. Based on fine tuning know-how, the best people proven through Kaggle and various international academic papers formed a task force to develop this open LLM model.

The Upstage 30B model's domination in the open-source LLM ranking marks a momentous turning point in Upstage's path towards establishing itself as a frontrunner in the rising "private AI" market worldwide.

Upstage 30B model highlights the potential of a compact and efficient LLM as a cost-effective alternative for companies looking to integrate generative AI into their business operations. While there is a growing demand for deploying state-of-the-art language models, many companies are reluctant to rely solely on ChatGPT or Bard for all their operations due to the inherent risk of data leakage.

To tackle this issue, private AI solutions are emerging as a new chapter in the generative AI revolution. These models are trained specifically on the company's private data, guaranteeing the security of sensitive information and reducing the possibility of producing inaccurate or misleading outputs. Industry bellwethers such as Apple, Walmart, Amazon, and JP Morgan have already banned their employees from using ChatGPT. Samsung Electronics is also taking steps to introduce its in-house language model to address security considerations.

With Upstage's small-but-powerful LLM, companies and institutions now have access to an advanced AI model without the worry of information leakage. This presents an opportunity for them to develop internal AI solutions by training the model using their own internal data and protocols. This enables employees to easily access and benefit from the company's knowledge management system. Another potential application is utilizing the AI solution to analyze sales data and create optimal marketing strategies.

In order to further enhance the performance of its model and meet the growing demands of domestic enterprises across various industries, Upstage has plans to train its model using an additional set of Korean language data. This proactive approach demonstrates Upstage's commitment to continuous improvement and its dedication to meeting the evolving needs of its clients in the local market.

Sung Kim, CEO of Upstage, said, "I am thrilled that Upstage’s AI capabilities dominated the global ranking.” He added, “Moving forward, Upstage will not only focus on enhancing the performance of the Korean language, but also strive to become a global leader in the emerging private AI market for businesses.”

← Back to Newsroom