Collaboration between KB and GPT-NL

GPT-NL is a Large Language Model (LLM) which has been developed as a responsible Dutch alternative to existing models. Language models are a type of artificial intelligence (AI) which have been trained to analyse human language, recognise patterns, and generate their own text. GPT-NL is an initiative by TNO, SURF and the NFI and has been trained on Snellius, the Dutch super computer.

The content board of GPT-NL is made up of parties who supply data to the language model. Together, they form the interest representation for the parties who have offered up their data to train GPT-NL.

The launching partners are the first parties who are performing feasibility studies with GPT-NL. These are scientific studies which test GPT-NL in practice. The results from these studies will be directly utilised in the development of the first version of GPT-NL.

As part of the content board and launching partner, the KB contributes to both the data set that trains the model and the language model's testing.

The KB finds it essential that there is a responsible Dutch alternative to the existing commercial providers of language models. GPT-NL is working on the development of an AI model for Dutch language and context, which will supply reliable information based on high quality data. As the KB, we happily contribute to GPT-NL's intention to an eminent and pluralistic data-ecosystem. 

A second reason for the KB's collaboration with GPT-NL is the fact both parties hold the use of legitimately obtained data in high regard. The KB has only supplied public domain data to GPT-NL: written heritage of which the copyright has expired. Like GPT-NL, the KB finds it important that sources are traceable and transparent, and that copyright owners get a fair place in the development of technology. A part of GPT-NL's income will be returned to copyright owners.

The feasibility study that the KB, National Library of the Netherlands is running with GPT-NL has the goal of providing more insight into the applicability of AI within the KB's collection search system. For this study, we are using a copyright free data set (works of more than 140 years old) from the KB platform Delpher.nl. As a specific case, we are developing an RAG-assistant with the goal of helping users to a suitable research introduction. Users of the assistant will no longer get thousands of possible Delpher search results, but a limited, relevant selection of sources.

The study's results will supply insights about the type of answers given by the language model to GPT-NL, while the KB will learn more about possible AI applications for its users.

GPT-NL has been developed for the business market. This means that Dutch people will not be able to use GPT-NL like consumers today can use ChatGPT, Le Chat, Claude or Gemini. In the future, Dutch organisations will be able to choose GPT-NL for specific tasks or use cases. For example: summarising texts, simplifying letters, or creating a manageable research design.

GPT-NL strives for a digitally autonomous Europe, in which responsible innovation is the norm. GPT-NL wants to show that it is possible to, in agreement with content suppliers, develop technology that works well for specific tasks. It's also a goal to learn more in the Netherlands and Europe about how this technology works, by being transparent in the development of GPT-NL about the data that goes into it and the choices that have been made.

Unlike many models available for the consumer market, GPT-NL is based on a more specific database, with data supplied by, among others, the KB, the Dutch Language Institute and the ANP. This also means the model has been developed for more specific tasks, such as summarising or simplifying text. Larger language models which have been based on larger data sets, such as ChatGPT, can also answer more generic questions. For example, it can suggest fun attractions on holiday or a simple recipe for a company of 5. Because GPT-NL has been developed for more specific applications in a Dutch context, it needs less data to train a model that still works well.

The funding conditions behind GPT-NL state that the running expenses need to be recouped, which means the language model can't simply be offered for free or under an open source license. GPT-NL will be made available through licenses. The license for business use will be offered at a market-based price. The license for research will be offered for a symbolic fee or for free.

There are parties who have supplied data and parties who are currently running feasibility studies with GPT-NL to test the model.

The data providers are, among others, the KB, the Nederlandsche Bank and the VNG. Besides the KB's feasibility studies, there are three more feasibility studies currently running with the Ministry of the Interior and for the website Overheid.nl. There is also a feasibility study being performed with TNO and the NFI.