Naver Cloud announced that it has added a tuning function to the No Code AI platform ‘Clova Studio’, maximizing the functionality of the Hyper Clova language model, enabling optimization for user purposes.
Clova Studio is a no-code AI platform that allows you to explore the applicability of AI and apply it to actual services even without development-related expertise. ▲’Playground’ where non-developers can experiment with AI based on text ▲’Explorer’ where they can browse and use AI that other users have worked on ▲’Forum’ where users can share opinions and Q&A with each other. It’s done.
Clova Studio, currently available as a closed beta service, has been continuously updating its features since its launch. The newly added tuning function learns user data and then optimizes and utilizes some of the Hyperclova language model parameters according to task type, language, and data.
Even without having their own AI technology and manpower, non-developers can easily and conveniently create the desired language model by uploading a certain amount of standardized datasets, making Clova Studio a more useful tool for startups that lack AI development capabilities.
The tuning function maximizes the potential of the Hyperclova language model. So far, we have only tried the in-context-learning method, which receives examples and instructions as input documents through prompts in the playground, understands their meaning, and performs tasks such as summarization, classification, and creation. In this case, the generation quality of the language model varies depending on the type of prompt configuration, and there is also a limit to the number of characters in the prompt. However, using the tuning function has the advantage of providing stable and superior performance compared to the in-context learning method and reducing the limit on the number of characters in the prompt and the dependence on prompts.
Additionally, through tuning, you can easily test whether the model has been learned appropriately for the user’s intent, and you can also use it as an application program interface (API) to actually apply it to the service.
In addition, it is noteworthy that costs and work time can be efficiently reduced. The Hyperclova language model can be applied to a variety of natural language processing (NLP) tasks at a lower cost than learning an entire large-scale language model. The performance reaches the level of existing large-scale language models that have learned tens of thousands of cases by learning only a few hundred to thousands of small datasets.
Through current tuning, a total of 6 NLP tasks can be performed: ▲document double classification ▲multiple classification ▲sentence summary ▲sentence generation ▲sentence correction ▲writing style conversion. Naver has been conducting pilot operations since June of this year and has used it extensively for NLP tasks such as conversation, summarization, translation, and classification, greatly improving service performance. For example, the AI care phone service ‘Clova Care Call’ improved its data inspection rate from 30% to 91% through tuning.
Meanwhile, the tuning function is only provided for the Korean model, but multilingual models such as English will be added, and the types of NLP tasks that can be performed and tuning techniques will continue to be expanded. Clova Studio can be used and applied through the Naver Cloud Platform.